9512.net
甜梦文库
当前位置:首页 >> >>

Supervised by Prof. Hozumi Tanaka


The Analysis of Japanese Relative Clauses

A DISSERTATION SUBMITTED TO THE DEPARTMENT OF INFORMATION ENGINEERING OF THE TOKYO INSTITUTE OF TECHNOLOGY IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF MASTER OF ENGINEERING

Supervised by Prof. Hozumi Tanaka

Timothy Baldwin February 1998

Contents
1 Introduction 1.1 Objectives and outline . . . . . . . . . . . . . . . . . . . 1.1.1 Statement of purpose of this research . . . . . . . 1.1.2 Methodological outline . . . . . . . . . . . . . . . 1.1.3 Applications of this research . . . . . . . . . . . . 1.2 A basic model of Japanese syntax . . . . . . . . . . . . . 1.3 De?nitions . . . . . . . . . . . . . . . . . . . . . . . . . . 1.3.1 Coordination, cosubordination and subordination 1.3.2 Displaceability . . . . . . . . . . . . . . . . . . . 1.4 Thesis overview . . . . . . . . . . . . . . . . . . . . . . . 2 Background 2.1 The structure of Japanese relative clauses 2.2 Past research . . . . . . . . . . . . . . . . 2.2.1 Descriptive accounts . . . . . . . . 2.2.2 Relativisation and thematisation . 2.3 Relative clause type de?nitions . . . . . . 2.3.1 Case-role gapping relative clauses . 2.3.2 Head restrictive relative clauses . . 2.3.3 Full clause-based idioms . . . . . . 2.4 Distribution of the relative clause types . 2.5 The full relative clause type hierarchy . . 3 Valency, Argument Types and Case 3.1 Complement/adjunct distinction . . 3.2 Argument status . . . . . . . . . . . 3.3 Case set . . . . . . . . . . . . . . . . 3.3.1 Core grammatical cases . . . 3.3.2 Other complement cases . . . 3.3.3 Local cases . . . . . . . . . . 3.3.4 Time cases . . . . . . . . . . 3.3.5 Oblique cases . . . . . . . . . 1 1 1 2 2 2 3 3 4 4 5 5 6 6 9 9 9 12 13 13 13 15 16 18 21 21 24 26 27 27 29 29 29 31 32 32

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

4 Verb class-based resolution 4.1 Verb class hierarchy . . . . . . . . . . . . . . . . . . . 4.1.1 Argument status and interpretation preference 4.1.2 Constructing the verb class hierarchy . . . . . 4.2 Physical movement . . . . . . . . . . . . . . . . . . . . 4.2.1 Distal movement . . . . . . . . . . . . . . . . . ii

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

CONTENTS 4.2.2 Proximal movement . . . . 4.2.3 Travelling . . . . . . . . . . Relational . . . . . . . . . . . . . . 4.3.1 Inter-personal relational . . 4.3.2 Generic relational . . . . . 4.3.3 Part-whole relational . . . . Valence variational . . . . . . . . . 4.4.1 Con?ated ergative . . . . . 4.4.2 Partitive . . . . . . . . . . . Other verb classes . . . . . . . . . 4.5.1 Copula . . . . . . . . . . . 4.5.2 Conjoining . . . . . . . . . 4.5.3 Quantative . . . . . . . . . 4.5.4 Existential . . . . . . . . . 4.5.5 Locational action . . . . . . 4.5.6 Tool-aided action . . . . . . Quotative . . . . . . . . . . . . . . 4.6.1 Subordinate clause gapping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

iii 32 34 34 35 36 36 38 38 39 40 40 40 40 41 42 43 44 44 46 46 46 47 47 48 49 49 50 50 52 52 53 53 54 54 55 57 57 58 58 59 59 60 61 61 61 62

4.3

4.4

4.5

4.6

5 Miscellaneous processing 5.1 Non-gapping expressions . . . . . . . . . . . . . . 5.1.1 The extraction of non-gapping expressions 5.2 Time-related adjuncts . . . . . . . . . . . . . . . 5.2.1 Temporal masking . . . . . . . . . . . . . 5.2.2 Time relative constructions . . . . . . . . 5.2.3 Temporal expressions . . . . . . . . . . . 5.2.4 Temporal vs. Durational interpretations . 5.3 Cardinal adjuncts . . . . . . . . . . . . . . . . . . 5.4 The default rule set . . . . . . . . . . . . . . . . 6 Lexical ambiguity 6.1 Verb lexical ambiguity . . . . . . . . . . . 6.2 Resolving verb lexical ambiguity . . . . . 6.2.1 Calculation of verb scores . . . . . 6.2.2 Complexity of in?ectional content 6.2.3 Evaluation of verb scoring . . . . . 6.3 Noun head lexical ambiguity . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

7 Extensions to the basic algorithm 7.1 Relative clause cosubordination . . . . . . . . 7.1.1 Processing of clause cosubordination . 7.1.2 Gap correspondences . . . . . . . . . . 7.1.3 The treatment of subordinate clauses . 7.2 Coordinated relative clauses . . . . . . . . . . 7.3 Noun head coordination . . . . . . . . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

8 Evaluation 8.1 Evaluation criteria . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.1.1 Baseline evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.2 Overall evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

iv 8.2.1 Evaluation 8.2.2 Evaluation 8.2.3 Evaluation 8.2.4 Evaluation Verb class-speci?c of ?xed expressions . . . . . . . of subordinate gapping . . . . . of clause cosubordination . . . . of inter-personal relational verbs evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

CONTENTS . . . . . . . . . . . . . . . . . . . . . . . . . 63 64 65 65 67 74 75 76 76 76 79 80 81 81 82 83 83 83

8.3

9 Conclusions 9.0.1 Future research . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A The Relative Clause Resolution System A.1 The system valency dictionary . . . . . . . . . . . . A.1.1 Valency frame types . . . . . . . . . . . . . . A.1.2 Valency frame type preferences . . . . . . . . A.1.3 Dictionary entry content . . . . . . . . . . . . A.2 NTT valency dictionary . . . . . . . . . . . . . . . . A.3 System valency dictionary extraction . . . . . . . . . A.3.1 Manual correction/modi?cation of the system A.3.2 Verb in?ectional analyser . . . . . . . . . . . A.3.3 Valency frame compatibility analyser . . . . . A.3.4 Semantic dictionary / lexical analysers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . valency dictionary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

List of Tables
4.1 4.2 6.1 8.1 8.2 8.3 8.4 8.5 8.6 8.7 8.8 8.9 8.10 8.11 8.12 8.13 8.14 8.15 8.16 8.17 8.18 8.19 8.20 8.21 8.22 8.23 8.24 Distribution of verb class frequencies . . . . . . . . . . . . . . . . . . . . . . . . . . . . Verb class frequency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Results for the verb scoring methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . Overall analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . Analysis of ?xed expressions . . . . . . . . . . . . . . . . . . . . Results of subordinate gapping analysis . . . . . . . . . . . . . Results of cosubordinated clause analysis . . . . . . . . . . . . Analysis of inter-personal relational verbs (without reciprocity) Analysis of inter-personal relational verbs (with reciprocity) . . Analysis of con?ated ergative verbs . . . . . . . . . . . . . . . . Analysis of Conjoining verbs . . . . . . . . . . . . . . . . . . . . Analysis of the copula . . . . . . . . . . . . . . . . . . . . . . . Analysis of distal movement verbs . . . . . . . . . . . . . . . . Analysis of empathy verbs . . . . . . . . . . . . . . . . . . . . . Analysis of excluding verbs . . . . . . . . . . . . . . . . . . . . Analysis of exisential verbs . . . . . . . . . . . . . . . . . . . . Analysis of experiential verbs . . . . . . . . . . . . . . . . . . . Analysis of generic relational verbs . . . . . . . . . . . . . . . . Analysis of idioms . . . . . . . . . . . . . . . . . . . . . . . . . Analysis of including verbs . . . . . . . . . . . . . . . . . . . . Analysis of locational action verbs . . . . . . . . . . . . . . . . Analysis of non-quantative verbs . . . . . . . . . . . . . . . . . Analysis of partitive verbs . . . . . . . . . . . . . . . . . . . . . Analysis of proximal movement verbs . . . . . . . . . . . . . . . Analysis of quotative verbs . . . . . . . . . . . . . . . . . . . . Analysis of tool-based action verbs . . . . . . . . . . . . . . . . Analysis of travelling verbs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32 33 55 62 64 64 65 66 66 67 67 68 68 68 69 69 69 70 70 70 71 71 72 72 72 73 73 83

A.1 Case valence dictionary statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

v

List of Figures
1.1 2.1 2.2 2.3 4.1 A valency frame model of Japanese syntax . . . . . . . . . . . . . . . . . . . . . . . . . The structure of Japanese relative clauses . . . . . . . . . . . . . . . . . . . . . . . . . The relative distributions of the proposed Japanese relative clause types . . . . . . . . The relative clause type hierarchy and associated sub-classi?cations . . . . . . . . . . . The full verb class hierarchy (Original verb classes indicated in bold, partitioning nodes capitalised) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 6 14 14 30 77

A.1 The full system hierarchy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

vi

Acknowledgements
First and foremost, I would like to express my heartfelt gratitude to Prof. Hozumi Tanaka in supervising me over the past two years, giving me the opportunity to come to Japan, and making every step of the journey to reach this point smooth and enjoyable. At the same time, Assoc. Prof. Takenobu Tokunaga has generously given of his time to help me take my ?rst shakey steps towards becoming a genuine researcher. Thanks go also to the long-su?ering residents of Room 807 and all other members of the Tanaka and Tokunaga laboratories, in being asked for opinions on reams of example sentences, getting me and my computer back on speaking terms on numerous occassions, and for doing it all with a genuine smile all the way. I doubt that this research would have progressed anywhere near the stage it has, without the assistance of the NTT Machine Translation group in providing access to their unsurpassed range of resources. In particular, I cannot thank Francis Bond enough for patiently reading papers in varying stages, sharing his work, and generally being a good friend throughout the past two years. I look forward to getting back to my de?niteness ‘roots’ with Francis now that I have this little fellow out of the way! And most importantly, I would like to thank family and friends for being patient and understanding while I learnt how to do things the hard way. I would like to say that I will devote more of myself, my time, and my energy to being a better husband, brother, son, and friend from now on, but realistically know that I am only going to turn around and plunge myself back into the cesspool of relative clauses and put you all through more of the same!

vii

Abbreviations
abl acc all cause com dat dim gen loc mutual neg -nml nom pass past pres prog qp quot ren top while Ablative Accusative Allative Causative Comitative Dative Familial Diminutive Genitive Locative Mutual Negative Nominaliser Nominative Passive Past tense Non-past tense Progressive Question Particle Quotative Ren’yo Topic While (nagara)

viii

Chapter 1

Introduction
1.1
1.1.1

Objectives and outline
Statement of purpose of this research

The purpose of this research is to classify Japanese relative clauses according to the nature of the semantic relationship between the relative clause body and the noun head. We consider the relative clause construal process as being both syntactically and semantically governed, and weighted pragmatically. By this is meant that for any given relative clause complex, there will be a well-de?ned set of semantic links which can be drawn between the clause body and head, based on both case-role interpretation and head restriction from the relative clause; these semantic links will each be constrained by such factors as sortal preferences on the head, in?ectional restrictions on the main verb, and case slot instantiation. In order to extract and label these semantic links, we propose a theory of case-role types and verb class correspondences, and assign a mini-rule set to each verb class for use by the relative clause resolution system. The job of the system is ?rst to syntactically and semantically determine the scope of the relation set, and then to apply pragmatic preferences to determine the plausibility of each such relation. The ?nal system output should consist of an ordered list of the ?nal interpretation candidates, with the highest ranking interpretation taken as the ?nal unmarked interpretation, barring further adjustment of the relative weights by external modules. To take an example relative clause complex of manzokusita gakusei “a satis?ed student”, the system should ?rst be able to determine the unacceptability of case-role gapping interpretations such as Direct Object, Indirect Object and Co-actor, syntactically from the valency frame of manzoku(-suru). At the same time, the semantics of the verb and noun head should lead to the disallowment of specialised head restrictive senses such as Inclusive and Exclusive, and adjunct interpretations of the locative and temporal types. This would leave the two candidates of Subject case slot gapping and relative clause-based head restriction, from which the system would be expected to correctly select the Subject gapping interpretation. An additional sub-purpose of this research is to construct a broad-coverage verb class hierarchy which can potentially be used to predict valency frame alternations/transformations (in the manner of Levin (1993)), and case-role correspondences. In this, the role of verb classes is twofold: (i) to model relations between case-roles within the valency frame, and (ii) to document the semantic type of the situation/action described by the verb. It is hoped that this classi?cation will have wider ranging applications to the analysis of discourse processes, and can be used both in extracting zero pronominal instances from Japanese text and describing the interaction of verb arguments within the discourse. Lastly, argument types will be introduced to rank case-roles according to contribution to the predicate sense, and apply heuristics with which to predict argument functionality. 1

2

CHAPTER 1. INTRODUCTION

1.1.2

Methodological outline

The methodology for analysing relative clauses devised in this research has been formulated so as to minimise use of extra-linguistic (i.e. pragmatic) resources in extracting the semantic relationship between the head noun and its associated relative clause. Importantly, we maximise reliance on syntactic and morphological analysis to this end, and use only limited semantic classi?cation techniques throughout the extraction process. While this may appear contradictory to the semantic motivation underlying the research, it is intended to demonstrate the intrinsic relationship between (syntactic) case marking and case-role identi?cation in Japanese, and the syntactic predictability of in?ection-based valence transformation. We thus do not necessarily wish to oppose ourselves to the semantic/pragmatic-founded work of authors such as Sato (1989), Matsumoto (1990, 1996, 1997), Neumann (1994), Kanzaki (1997), and Kanzaki and Isahara (1997), and willingly admit that pragmatic and discourse-related factors can override syntactic and semantic preferences. However, given that the current work is focused at relative clause processing in isolation, without access to discourse and situational context, this design decision to predict unmarked relative clause construal would appear unavoidable. At the same time, all e?ort has been made to develop a modular system which can hypothetically interface with a discourse processor to determine the most plausible interpretation from the ranked candidates outputted by the proposed system, and potentially reweight them according to such factors as contextual salience and recency, as well as sortal preferences for the head noun in performing given case-roles.

1.1.3

Applications of this research

Japanese involves the extensive usage of relative clause constructions. Given that approximately 50% of sentences contain one or more relative clauses, and that there are approximately two relative clauses for every three sentences in an average text,1 the accurate analysis of relative clauses would seem to be a vital element of any overall system. This in itself would seem to justify any attempt to provide an accurate, robust algorithm for analysing Japanese relative clauses. Theoretically, the method presented herein should be congruent with any NLP system, given its self-containing nature and relatively low-cost interface. Particular applications which would bene?t most from accurate relative clause analysis would be Japanese language understanding tasks, including information retrieval and document extraction, and machine translation methods from Japanese to a Western language such as English2 . Admittedly, however, in order to o?er tangible results to the machine translation community, further work is required to transfer the analysis types proposed in this research over to translation equivalents in the target language.

1.2

A basic model of Japanese syntax

Case and Valency provide valuable tools in describing Japanese syntax, and are called upon frequently throughout this paper. The predicate is taken to be the nucleus of the clause, and relies on Valency to de?ne the range and type of “case slots” subsumed by that predicate, according to the predicate modality. Each case slot is associated with a distinct “case-role” and “case marking” type. The case-role (aka. theta role, Case, semantic role, etc.) is an account of the semantic contribution of that case slot to the semantics of the predicate, although it is de?ned based on grammatical position
The ratios given are based on an analysis of the EDR corpus (EDR 1995). Of the total of 201340 sentences, 99673 contained one or more relative clauses, and the total number of relative clauses was 133655. 2 The applicability of the system and general methods utilised therein to language generation tasks is a matter for further research, although I suggest that the verb class hierarchy can potentially contribute to any Japanese language task relying on the notions of Case and Valency.
1

1.3. DEFINITIONS

3

Case ?ller-case markerCase-role Case slot 1

Case ?ller-case markerCase-role Case slot 2

···

Predicate

Figure 1.1: A valency frame model of Japanese syntax

for the core case-role set. To take an English example, The party was held at David’s house contains a Subject and a Locative, in the form of the party and at David’s house, respectively. Other examples of case-roles are Direct Object, Co-actor, Perlative, and Instrument. Japanese marks each argument (case ?ller) with morphological case, in the form of a phrase-?nal case marker. Thus, in the Japanese equivalent for the above English example, p? ati-ha D? ebiddo no iede okonawareta (party-topDavid gen house-locwas held), the Subject p? ati is marked with the topic marker and the Locative D? ebiddo no ie is marked with the locative case marker. These case-role/case marker tuples make up the content of each case slot. That D? ebiddo no ie both performs the Locative case-role and is marked in the locative case is not entirely incidental, and peripheral case-roles commonly coincide with their canonical case marking type (another example of this is the Comitative case-role and comitative case marking), but that is not to say that this is either a necessary or su?cient condition. The case-role schema and case marking patterns should thus be considered as orthogonal issues, and the reader should bear in mind that in recovering the case-role of a given case slot, word order and the semantic content of the case ?ller also play integral roles.3 It is also true that a number of the proposed case-roles are associated with a unique case slot (in particular the Durational and Instrument case-roles), and that the core roles have a default case marker, but in no sense does the reverse apply, due to the con?ation of case marking. To take an extreme case, the dative case marker (ni) can mark almost any case-role and provides minimal indication of the particular case-role of that case slot. For this reason of case marking variation, also, we consider case-roles and case marking as separate issues. The valency frame for a given verb and verb sense is made up of individual case slots, with the scope of case marking for each slot often being plural. We will refer to the case marking paradigm for all case slots, considered in isolation of case-role correspondence, as the case frame. Case frames can thus be derived trivially from valency frames by discarding case-role information.

1.3
1.3.1

De?nitions
Coordination, cosubordination and subordination

Clausal relations provide a valuable mechanism when analysing complex relative clauses, and are discussed variously throughout this thesis. In describing clausal relations, we apply a parameterised trichotomy comprising the “subordination”, “cosubordination” and “coordination” types (Foley and Van Valin 1984; Van Valin 1984). Two parameters are used to di?erentiate these three types: dependence and embedding. Dependence is form-based, and relates to whether the clause in question is syntactically dependent on surrounding clauses (either for operators or distributionally), or alternatively can stand alone as a complete sentence. Embedding, on the other hand, describes whether the clause is encapsulated within/functions as part of another clause, or is ‘complete and distinct’ (Van Valin 1984:542); herein, embedded clauses will be used to refer to quoted clauses marked with the quotative case marker.
3

See (Blake 1994:13-8) for a discussion of this process in the context of a variety of language types.

4

CHAPTER 1. INTRODUCTION

Combining these two independent parameters, we in fact produce a four-way taxonomy, of which only the three classes given above are relevant for our purposes. a. b. c. d. Coordination Parenthesis Cosubordination Subordination [?dependent, [?dependent, [+dependent, [+dependent, ?embedded] +embedded] ?embedded] +embedded]

To clarify this distinction, [David ran] [and Peter rode his bike] is an instance of clause coordination, [David ran] [because Peter was riding his bike] is an instance of clause cosubordination (distributional dependence), and [David ran to allow Peter [to ride his bike]] is an instance of clause subordination.

1.3.2

Displaceability

Displaceability is a statement of the potential for a given case slot to be case-role gapped from a relative clause context, a notion original to this research. We wish to claim that it is possible to make an a priori judgement on the displaceability of a given case slot, independent of case ?ller type, and that if provided with a case ?ller which can be expressed in a displaceable case slot for a matrix clause instance, that case ?ller can also be gapped to become the noun head of a relative clause. In the case of the verb nar(-u) “to become”, for example, and the “Subject-nomDirect.Object-dat” valency frame, the Subject case slot is displaceable, but the Direct Object is not: (1) a. sy? ob? osi-ni natta kare ?reman-dat became he (lit.) “he, who became a ?reman” * kare-ga natta sy? ob? osi he-nom became ?reman “the ?reman he became”

b.

1.4

Thesis overview

In Chapter 2, we de?ne what is meant by “relative clause”, and contextualise this research by giving a brief description of past treatments. We then describe the relative clause types around which this relative clause analysis pivots. Chapter 3 is an introduction to argument status, and their contribution to the proposed case-role schema. The full range of case-roles is illustrated with examples and descriptions of their synactic and semantic behaviour. In Chapter 4, we bring together both argument status and the case-role schema in detailing the verb class hierarchy proposed in this research, based upon which the modular rule sets applied in the resolution process are derived. Chapter 5 is an account of peripheral case-roles, and their interface to the verb class rule sets. Chapter 6 details the methods used to resolve lexical ambiguity both on the noun head and main verb, and again refers back to their relation to the rule sets. Chapter 7 describes inter-clausal and intra-noun head e?ects and their relevance to relative clause resolution, while Chapter 8 contains details of how the resolution system functions, what its parts are, and where they originate from. In Chapter 9, a thorough evaluation of the resolution system on di?erent data sets is presented, including discussion of unexpected weaknesses in the system, before concluding the thesis in Chapter 10.

Chapter 2

Background
2.1 The structure of Japanese relative clauses

The general syntactic structure of Japanese relative clause complexes is given in ?gure 2.1 below. The noun phrase (NP) head is modi?ed prenominally by a verb phrase (VP) clause body in comprising the overall noun phrase. We will refer to the NP head of the relative clause as the ‘noun head’ or ‘head’ throughout this paper, and the overall VP NP head-modi?ed complex as a ‘relative clause complex’. The current research is aimed at verb-based relative clauses, but relative clauses incorporating adjectives and adjectival nouns are theoretically included under the generic category of VP, with the proviso that a copular connective (na) is required between the clause body and head in the case of adjectival nouns. Despite the structural parallelism that exists between verb-based and adjectivebased relative clauses, however, it is suggested that there are certain semantic peculiarities displayed by adjectives which are not seen for verb-based relative clauses, drawing from the essentially stative nature of adjectives. Evidence of this can be seen in the work of Kanzaki (1997) and Kanzaki and Isahara (1997), but the absolute correlation between the two relative clause types is left as a moot question. Semantically, the noun head and clause body stem verb cannot include clause-level grammatical constructions or grammatical markers which are unable to exist as independent NPs. This leads to the preclusion of the following types of VP NP constructions from our de?nition of relative clauses:1 a. Formal nouns/postadnominals (Martin 1975:664-740) such as noti “after”, tame “for (the purpose of)” b? ai “case/circumstance”. These act as discourse/clause-level markers and unambiguously collocate with a clause body, deictic marker or noun speci?er. b. ‘Relational’ verb stems2 , as taken in the Hallidayean sense (see Halliday (1994:119-38)). This includes constructs of the type to-iu “called”, ni-kansuru “concerning” and ni-taisuru “against/regarding”. It is important to note here that our de?nition of ‘relative clause’ for Japanese includes both NP complexes that involve case-role gapping (an adaptation of the traditional de?nition of relative clauses), and those for which the relative clause body simply restricts/exempli?es the noun head.3 That is, our use of the term “relative clause” corresponds to Matsumoto’s “noun-modifying construction” (1996), and di?ers signi?cantly from the restricted Transformational Grammar sense of the word, which corresponds to the precept of ‘case-role gapping’ relative clausehood in our framework. Admittedly, this appears to go beyond the bounds of the standard sense of relative clausehood, but the terminology is intended to re?ect the syntactic parallelism that exists between these two relative
1 There is limited scope to apply the methods described here to the processing of relative clause-type constructions produced with these operators, although this is left as a matter for future research. 2 Referred to as “phrasal postpositions” by Martin (1975). 3 Kameyama (1995), likewise, uses the term ‘relative clause’ in this wider sense.

5

6

CHAPTER 2. BACKGROUND
RELATIVE CLAUSE CONSTRUCTION

manzoku-sita yuza NP

S manzoku-sita
RELATIVE CLAUSE BODY

NP yuza
CLAUSE HEAD

Figure 2.1: The structure of Japanese relative clauses

clause types. Additionally, while continual reference will be made to the notion of ‘gapping’, this is intended in a semantic case-role based sense, and we wish to distance ourselves from the research on syntactic gapping and movement that exists in the Chomskian literature (Nakau 1971; Okutsu 1974; Inoue 1976; Huddleston 1976; Shibatani 1978; Radford 1981). Clearly, similarities exist between the Chomskian treatment of gapping and its correlation to deep structure transformation, but, for our purposes, we view relative clause construal as a ‘focusing’ or ‘aboutness’ (Kuno 1976; Saito 1985; Kuno 1987) process rather than movement.

2.2
2.2.1

Past research
Descriptive accounts

Traditional descriptions of Japanese relative clauses have divided them into two main disjunctive categories, based on whether the noun head can be reinserted into the relative clause body to compose a matrix sentence (Martin 1975; Teramura 1975–78). Perhaps the most famous such account is that proposed by Teramura (1970, 1975–78, 1978, 1981, 1980), in which he describes the semantic relationship between the relative clause body and noun head either as the uchi no kankei “inner relationship” type or the soto no kankei “outer relationship” type, corresponding to the “sentence-insertable” and head noun content-supplementing sense designations, respectively. Following this typology, [ manzoku-sita ] gakusei “a satis?ed student” is an inner relationship relative clause, and [ manzoku-sita ] wake “a reason for feeling satis?ed” is an outer relationship relative clause. On close observation, Teramura’s schema attempts to classify inner relationship relative clauses as being syntactically-de?ned and outer relationship relative clauses as being semantically-de?ned, a point which is convincingly refuted by Matsumoto (1997) in citing cases where pragmatics in?uence interpretation of inner relationship relative clauses, such as:

2.2. PAST RESEARCH (1) a. Donarudo Toranpu-ga katta ] mise Donald Trump-nom bought shop “the store (which) Donald Trump bought” [ Tomo-tyan -ga katta ] mise Tomo-dim-nom bought shop “the store (at which) little Tomo bought ( [

7

b.

)” (from (Matsumoto 1997:43))

More recently, Saito (1985) reviewed the syntactic licencing of relative clauses, and suggested that syntax should account for the formation but not interpretation of inner relationship relative clauses. By way of relaxing the syntactic-dependence of inner-relationship-type relative clauses, it becomes possible to consider both inner and outer relationship type clauses in the same semantic light, and their equal susceptibility to pragmatic e?ects. There is a further di?culty in defending the integrity of a Teramura-style proposal, however. Whereas the demarkation between inner and outer relationship relative clauses is portrayed as a distinct one, there are cases of relative clause complexes which seem to ?t into both categories. Teramura recognises this shortcoming of his categorisation, and makes explicit mention of “truncated” constructions that straddle the divide between the two relationship types. An example of a truncated construction is: (2) Tar? o-ga p? ati-ni deta ] Taro-nom party-dat attended “Taro’s reason for attending the party” [ riy? u reason

Within the unique indicated interpretation, the relative clause can be viewed both as exemplifying riy? u, and as accommodating a gap for the noun head in the Purpose case slot: (3) Tar? o-ga (nan no) riy? u-de p? ati-ni deta Taro-nom (what gen) reason-loc party-dat attended watasi-ni wakaranai. I-dat not understand “I do not understand [Taro’s reason for attending the party].” [[ ] no ka ha -nmlqptop ]

While this structural/analytical ambiguity seems pertinent for unmarked content noun heads, however, scope di?erences arise for internally modi?ed noun heads, suggesting that the matter of truncation is not what it appears, and that the analytical ambiguity observed above is coincidental rather than terminal to Teramura’s classi?cation. (4) Tar? o-ga p? ati-ni deta ] hont? o no riy? u Taro-nom party-dat attended real gen reason “Taro’s real reason for attending the party” ? Tar? o-ga hont? o no riy? u-de p? ati-ni deta. Taro-nom real gen reason-loc party-dat attended “Taro attended the party for a real reason.” (= (4)) [

(5)

Similar observations can be made for time relative expressions (see Section 5.2.2), the other main source of truncation and contributor of claimed counterexamples to the inner/outer relationship dichotomy. In Matsumoto’s pragmatically-founded account of relative clauses, complexes such as (3) are interpreted as evidence for the inherent “framing” process involved in relative clause construal, through the predicate (“predicate framing”) or the noun head (“nominal framing”), or both, with truncated relative clauses involving mutual predicate/nominal framing. Inner relationship relative clauses form a component of the predicate frame modifying type, and outer relationship relative clauses fall into the nominal and mutual predicate/nominal framing modifying types.

8

CHAPTER 2. BACKGROUND

Matsumoto’s frame semantic approach seems to provide valuable discrimintative power to the subdivision of Teramura’s outer relationship relative clauses, both in terms of the source of the framing and additionally through the sub-type of the framing process. It lacks in credibility, however, in the semantic tenability of determining the relevant frame, identifying the available slots, and analysing compatibility between target arguments and the candidate slot set. Indeed, this ?rst step of frame determination seems to be problematic in a cognitive context, even, for non-situational frames. For the well-de?ned experiential example of “eating”, such as that described by Matsumoto for the predicate tabe(-ru) (Matsumoto 1997:61-63), the roles and frame are easily recoverable, but for a more abstract predicate such as “falling value”, the location of an appropriate frame and role set would appear more confused and the scope of role matching less apparent. Additionally, Matsumoto’s proposed methodology of evaluating case-role compatibility is not able to account for cases of bounded gapping. Perhaps more serious, however, is that the discriminative nature of Matsumoto’s expanded set of relative clause types produces its own demarkation problems. That is, as compared to the claimed di?culty in accounting for “truncated” relative clauses within Teramura’s schema, Matsumoto’s framework has inherent di?culties in delineating nominal framing (“NH-type”) and mutual predicate/nominal framing (“CNH-type”) relative clauses, and likewise nominal/predicate framing and predicate framing (“CH-type”) relative clauses. To take an example from (Matsumoto 1997:159), nioi in (6) is suggested as both participating in the frame evoked by the modifying clause (predicate framing) and evoking its own “relational” frame to describe the cause or source of the smell (nominal framing). As such, (6) is described as being of the CNH-type. (6) [ sakana-o yaku ] nioi ?sh-acc grill smell “the smell of grilling ?sh” [ sakana-o yaku ] kemuri ?sh-acc grill smoke “smoke from grilling ?sh” sakana-o yaku ] oto ?sh-acc grill sound “the sound of grilling ?sh”

This begs the question as to the status of (7) and (8). (7)

(8)

? [

One addition body of research worthy of comment is the semantically-based work of Sato (1989) in proposing an expanded set of relative clause types. For all intensive purposes, Sato retains Teramura’s inner relationship type as his kaku y? oso-gata “case element-type”, and partitions the outer relationship type into three subclasses, according to the nature of head modi?cation. However, he then goes on to propose a ?fth kansetsu gentei-gata “indirect restrictive-type”, in which the noun head ‘restricts a given case element in the relative clause by way of the genitive connective’ (Sato 1989:9 – my translation). He cites examples of this e?ect such as (9), which corresponds to the genitively-connected matrix clause given in (10). (9) tatemono-ga kowasareta ] tosi building-nom were destroyed city “the city (where) buildings were destroyed” (sono) tosi no tatemono-ga kowasareta (that) city gen building-nom were destroyed “(the/that) city’s buildings were destroyed” [

(10)

Further discussion of this “indirect restrictive-type” will be made in de?ning bound gapping relative clauses (see below).

2.3. RELATIVE CLAUSE TYPE DEFINITIONS

9

2.2.2

Relativisation and thematisation

One issue which has permeated the best part of (gapping) relative clause research, is the relationship between relativisation and thematisation. The ?rst mention of this matter can be found in (Kuno 1973a:254-5), where Kuno proposes that relativisation correlates to deletion of the “theme” (topic) of the relative clause. A partial motivation for this analysis is the realisation that (true) topics cannot occur within relative clauses4 , and additionally that relativisation is basically a focusing phenomenon, similar to thematisation; there are also close syntactic parallels between the two processes, such as the deletability of case marking, and argument type compatibilities (both can involve any of adverbial clauses, complex noun phrases and sentential Subjects). Despite the immediate problems inherent in Kuno’s proposal (Muraki 1970; Kuno 1973a; McCawley 1976; Matsumoto 1997), it does o?er an attractive guiding principle for di?erentiating between inner and outer relationship relative clauses, as most inner relationship relative clauses are sense-compatible with their topicalised noun head matrix equivalent, and outer relationship relative clause noun heads are almost exclusively not available to matrix topicalisation. We thus apply the topicalisation test implicitly as a litmus test to gauge the nature of relative clauses, but make no claims as to the equivalence or semantic coincidence of topicalisation and relativisation.

2.3

Relative clause type de?nitions

This research divides relative clauses into two major types, along the same lines as the traditional Teramura treatment, but adds a sub-type to the inner relationship type, by way of ‘bound’ relative clauses. Despite the obvious similarities to Teramura, we position ourselves very much away from the syntactic motivations of Teramura, and towards the semantic foundations of Matsumoto, Sato and Neumann.

2.3.1

Case-role gapping relative clauses

As previously described, “gapping” is used in a semantic sense to refer to the displacement of a case slot to the noun head position, in which respect case-role gapping relative clauses are distinctly removed from their origins in Teramura’s inner relationship type. Within the “case-role gapping” class, we go on to de?ne “bound” relative clauses, drawing on Sato’s indirect restrictive-type. Formally, case-role gapping is de?ned as the process whereby the noun head can be considered to ful?ll a unique well-de?ned case-role subsumed within the relative clause body, for a given interpretation. That is not to say that all relative clauses have a unique interpretation, but rather that, given a particular interpretation, the case-role gap will be linked to a unique case-role within the relative clause. Considering example (11) below, the gakusei noun head can be seen to perform the role of the Subject of the main verb, manzoku-sita. (11) manzoku-sita ] gakusei was satis?ed gakusei “a satis?ed student” (lit. a student who is satis?ed) [ sy? okai-sita ] hito introduced person a. “the person (who) introduced ( )” b. “the person (who) ( ) introduced” c. “the person (to whom) ( ) introduced ( [

(12)

)”

In contrast, the identity of the case-role gap for (12) is ambiguous between the Subject, Direct Object and Indirect Object case slots, but simultaneous gapping from these three case slots cannot occur.
4

It is possible to have case slots marked with the topic marker, in non-topic contrastive usages.

10

CHAPTER 2. BACKGROUND

That is, it is not possible to have an interpretation of the type the person (who) introduced (self ) ( ) or the person (who) introduced (self ) (to) (self ) without overt re?exive pronoun instances within the relative clause body. This leads to mutual exclusivity of case-role gapping between interpretations (12a), (12b) and (12c). Even in the case of the personal re?exive pronoun zibun occupying a case slot within the relative clause, coindexing occurs clause-internally through case-role gaps/zero pronominal case-roles, rather than directly with the noun head. (13) φi zibuni -o sy? okai-sita ] hito Sbj self-acc introduced person a. “the personi (who) introduced him/herselfi ” b. “the person to whom ( ) i introduced him/herselfi ” [

For (13), this correlates to zibun being coindexed with the Subject case-role, which in turn corresponds to either the case-role gap (cf. (13a)) or some other independent, zero-pronominal discourse participant (cf. (13b)). One feature of case-role gapping relative clauses is that whereas the case-role gap is de?ned uniquely for a given interpretation, the identity of the case slot from which gapping has occurred is not marked either as a trace within the relative clause, or as a relative pronoun-type marker. Moreover, there appear to be few restrictions on case-role positions from which gapping can occur, and when restrictions are found, they tend to be localised to that case-role in the given valency frame. Indeed, the main source of “restriction” is semantic, and derives from local sortal preferences de?ned through case frames.5 Returning to (12) above, any case-role ambiguity observable in the relative clause complex is removed in the matrix clause counterparts corresponding to the respective interpretations: (14) a. (sono) hito-ga sy? okai-sita that person-nom introduced “that person introduced ( )” (sono) hito-o sy? okai-sita that person-acc introduced “( ) introduced that person” (sono) hito-ni sy? okai-sita that person-dat introduced “( ) introduced ( ) (to) that person”

b.

c.

Bound relative clauses
Bound relative clauses are essentially Sato’s indirect restrictive relative clause type, except that we consider the process to be semantically governed, and extend Sato’s treatment to consider non-Subject case-role binding. The ‘binding’ facet of this relative clause type comes from its behaviour in anchoring a single attribute/‘part’ within the relative clause body. Note that the uniqueness of the gapped caserole seen for case-role gapping relative clauses, applies here for the bound case slot. The basic binding function can be realised in a matrix clause context either as the topic or by genitive linkage to the bound case-role argument. (15) a. [ 10-gatu-kara singakki-ga hazimatta ] pekin-daigaku October-abl new term-nom began Beijing University “Beijing University, which began its new term in October”

The actual process to determine which case slots case-role gapping can occur from, is largely left as a matter for future research.

5

2.3. RELATIVE CLAUSE TYPE DEFINITIONS b. 10-gatu-kara pekin-daigaku no singakki-ga hazimatta October-abl Beijing University gen new term-nom began “the Beijing University new term began in October” pekin-daigaku-wa 10-gatu-kara singakki-ga hazimatta Beijing University-top October-abl new term-nom began “the Beijing University new term began in October”

11

c.

In (15a), pekin-daigaku binds the instantiated Subject case slot, a fact which is recoverable through the head-argument genitive linkage in the Subject position for (15b), and topic realisation in (15c). For comparison, (16) is an example of a Bounded Direct Object relative clause, although ambiguity exists with a canonical Subject gapping sense. (16) hon-o katta ] sakka book-acc bought author a. “the author who bought the book” b. “the author, (whose) book ( ) bought” [

For complement case slots, binding is distinguishable from case-role gapping through the surface instantiation of the case-role in question. Naturally, the bound case slot must be instantiated for bound relative clauses (whether complement or otherwise), and the gapped case slot must be vacant for caserole gapping. Hence simple knowledge of the bound case-role de?nes the relative clause construal type for complement case slots, due to their uniqueness of surface form. However, the repeatability of adjunct case slots (see Section 3.1) means that simple adjunct instantiation does not necessarily lead to case-role gapping incompatibility. For the Locative case slot, multiple instantiation produces ‘inner’ and ‘outer’ Locative positions, and gapping can occur from either in the presence of the opposing mate.6 Likewise, for the Temporal case slot, case-role gapping can occur for a temporally ground relative clause given that ‘temporal masking’ constraints are met (see Section 5.2.1), as seen in (17) below. (17) [ 17-niti-ni kaigi-ga hazimatta ] 17th-dat meeting-nom started “the time the meeting started on the 17th” zikan time

This leads to total incompatibility of binding with the Temporal case slot, and compatibility only with ‘dangling local relational nouns’ for the Perlative case slot. Local relational nouns describe a location relative to some ?xed point, such as mae “front”, ue “top” or migi “right”. Dangling local relational nouns are produced when a bare local relational noun occupies a case slot. An example of a dangling local relational noun with the Perlative case slot is given in (18), which is contrasted with the ungrammatical dangling Allative usage in (19). (18) [ hik? oki-ga ue-o tonda ] sima plane-nom top-acc ?ew island “the island which the plane ?ew over” hik? oki-ga migi-ni tonda ] sima plane-nom right-dat ?ew island “the island which the plane ?ew to the right of” (intended)

(19)

*[

Dangling local relational nouns are also produced, however, for passives, from the Perlative or Direct Object case slots, suggesting that simple detection of a dangling local relational noun is not su?cient in itself to produce a bound relative clause.
For locational action verbs, the inability to reproduce inner locative case-role gapping to complement an outer locative instance within the relative clause, is a symptom not of the inconsistency of inner locative gapping itself, but of the impossibility to maintain the outer locative in an outer position without inner locative collocation.
6

12 (20) sih? o-o yama-ni kakomareta all four sides-acc mountain-dat is surrounded “a hut surrounded on all four sides by mountains” [ p? ezi-ga otiteiru ] page-nom is missing “a book with missing pages” hon book [ ]

CHAPTER 2. BACKGROUND koya hut

(21)

(22)

hon no p? ezi-ga otiteiru. book gen page-nom is missing “The pages of the book are missing.” b. hon-kara p? ezi-ga otiteiru. book-abl page-nom is missing “The book has missing pages.”

a.

2.3.2

Head restrictive relative clauses

Head restrictive relative clauses display identical surface syntactic structure to case-role gapping clauses. In the case of head restrictive relative clauses, however, the head represents a consequence, condition, requisite, simultaneous event, etc. of the modifying clause (Matsumoto 1997:103-130), or is simply restricted by the semantic content of the relative clause. Examples of head restrictive relative clauses are: (23) [ katu ] isi wins will “the will to win” [ au ] kikkake meets chance “a chance to meet ( [

(24)

)”

(25)

sakana-o yaku ] kemuri ?sh-acc grills smoke “smoke from grilling ?sh” gakk? o-ni itta ] kaeri school-dat went return “on the way back from school”

(26)

[

(27)

p? ati-ga hirakareru ] zenzitu party-nom is held the day before “the day before the party was/will be held”

[

In (23) and (24), the respective heads of isi and kikkake are simply restricted by their modifying relative clause, whereas kemuri in (25) is an inferrable consequence of the event described by the associated relative clause. (26) and (27), on the other hand, are headed by nouns which are relational in the given settings. For kaeri, some locative source from which it is possible to return, must be supplied to the noun head to ground the directional relation, and similarly, zenzitu requires a locative reference point before the day before can be calculated. We return to discussion of (27) and its relation to Temporal gapping relative clauses in Section 5.2.2. One test for head restrictive relative clauses is their incompatibility with a topicalised matrix clause context.

2.4. DISTRIBUTION OF THE RELATIVE CLAUSE TYPES

13

2.3.3

Full clause-based idioms

Full clause-based idioms are NP idioms constructed as a relative clause complex. Notice that the full complex construes the idiomatic sense, in which way full clause-based idioms are distinguished from ?xed expression sense in the relative clause. Due to the idiomatic unit nature of full clause-based idioms, it seems meaningless to attempt to analyse constructions of this type by way of the proposed case-slot gapping/head restrictive relative clause dichotomy. They are thus excluded from the classi?cation process, and simply marked as idioms on detection. Examples of full clause-based idioms are: (28) mite minu ] huri to look to not look pretence “close one eyes on ( )” Tar? o-to awaseru ] kao ] ga nai Taro-com put together face nom have not “I am ashamed [to show my face in front of Taro].” [[ [

(29)

For (29), it could be argued that kao has been case-role gapped from the Direct Object case slot, but this seems to gain no bene?t in terms of idiom analysis/comprehension.

2.4

Distribution of the relative clause types

In any discussion of relative clause types, it is worth considering the relative distribution of each type, and hence their relative importance in any analysis attempt (Figure 2.2). This calculation of distribution was carried out on the annotated relative clause corpus sourced during system evaluation, which originates from the EDR corpus (EDR 1995). Given the signi?cant size of the relative clause corpus (4615 disambiguated relative clause instances), we suggest that the indicated ?gures are indicative of the relative proportions the various types could be expected to occupy for similar-styled texts.7 From the ?gure, it is clear that case-role gapping relative clauses far outweigh head restrictive clauses (85% vs. 14%), and hence we feel justi?ed in having focused to this point on the analysis of case-role gapping relative clauses. The 1% ?gure for bound relative clauses represents a signi?cant proportion, although again, emphasis clearly lies on full case-role gapping.

2.5

The full relative clause type hierarchy

Figure 2.3 provides an account of the full relative clause type hierarchy, and the sub-classi?cations associated with each relative clause type. The heavy weighting of sub-classi?cations for case-role gapping is noticeable, and a fuller analysis of semantic linking types for head restrictive relative clauses would undoubtedly help o?set this unbalance.

The major component of the sentences contained in the EDR corpus are from newspaper articles, with lesser numbers of sentences from scienti?c texts and other miscellaneous sources.

7

14

CHAPTER 2. BACKGROUND

Bound
(1%)

Head restrictive
(14%)

Full-clause based idioms
(1%)

Case-role gapping
(84%)

Figure 2.2: The relative distributions of the proposed Japanese relative clause types

Case-role gapping

Head restrictive

Full idiom

Subject Direct Object Indirect Object Co-actor Passive Agent Causee Comitative Co-patient Object allative Beneficiary Object allative Locative Ablative Allative Perlative Durational Temporal Instrumental Cardinal

Bound gapping Bound Subject Bound Direct Object Bound Indirect Object Bound Co-actor Bound Perlative

Content Time Relative Exclusive Inclusive

Figure 2.3: The relative clause type hierarchy and associated sub-classi?cations

Chapter 3

Valency, Argument Types and Case
In describing case-role gapping, we clearly require a set of case-roles powerful and wide-ranging enough to label all case slots in all valency frames. Additionally, in order to uniquely label each case slot contained in a given valency frame, the granularity of the case-role designation must be ?ne enough to account for semantic di?erences between case slots. As an orthogonal issue, we introduce the concept of ‘argument status’, an expansion of the conventional complement/adjunct distinction. Argument status is used to predict argument obligatoriness, invaluable in distinguishing between zero pronominal and unrealised case slots, to introduce preferences between valency frames according to argument content, to gauge and generalise semantic consistency of usage, and to weight case-role gapping interpretations. This ?nal concern relates to which case-roles are preferred in a semantically neutral relative clause context of the type: (1) [ taberu eat ] X

That is, assuming the semantics of the noun head X are inaccessible, what case-role interpretation (ignoring head restrictive relative clause interpretations for the time being) would be most likely? Inevitably, (obligatory) complements are preferred over optional complements, with the particular ranking of preference of complement case-roles often relating back to topicality/accessibility hierarchies (Keenan and Comrie 1977; Inoue 1976) such that, in the case of (1), the Subject case-role would be preferred over the Direct Object, followed by the Locative case-role, and possibly the Instrumental case-role. At the same time, however, the scope of case-roles available in a given clausal context is clearly constrained by the predicate valency frame, and it would not be possible to have an Indirect Object gapping interpretation, for example. This illustrates the lowest level preference we can draw upon, de?ned by case-role immediacy and availability. Working against any such default ranking are a?nities for particular case-roles to take certain argument types in unmarked usages, and the potential for more specialised preferences for each caserole when in the context of a given valency frame and predicate sense. This ?rst issue can be seen with the converse form of (1), in which the predicate is unspeci?ed for a given lexical head: (2) [ PRED ] basyo place

Here, the most likely case-role mapping would be onto the Locative case slot, assuming locative adjunct compatibility for the predicate. Failing any particular semantic correspondence to a predicateindependent case-role type, however, default preferences of the type seen above would apply. This exempli?es the opposite end of the scale of argument status, and adjuncy. A third factor in this process is idiomatic usage, and overrides absolutely any local preferences 15

16

CHAPTER 3. VALENCY, ARGUMENT TYPES AND CASE

arising from case-role immediacy and adjuncy. For example, for the predicate utu “to strike” with an unmarked head X , the most accessible case-role would be the Subject, and equally for the noun head denp? o “telegram” with an unmarked predicate PRED, adjuncy would not have any e?ect, but the moment these two are combined as [ utu ] denp? o “the telegram ( ) sent”, the Direct Object case-role overrules any preferences to take absolute preference. Equally, an unmarked instance of the predicate a(-u) “to coincide” would subsume a Co-actor case-role, but when collocating with the argument keisan-ga “calculation-nom”, the valency frame is reduced to a single topic position. This chapter discusses an argument status hierarchy for use in predicting the syntactic interplay between arguments and predicates, and a case-role schema for documenting the semantic aspect of argument-predicate linkage.

3.1

Complement/adjunct distinction

The delineation between complements and adjuncts remains a contentious issue in valency research, with the most commonly occurring disagreement relating to the treatment of non-obligatory complements. Unlike adjuncts, which are by de?nition optional1 , complements are not necessarily obligatory. That is, all obligatory constituents are complements, and all adjuncts are omissible, but an optional argument can be either an adjunct or a complement. In an attempt to defuse the controversy surrounding the classi?cation process, we introduce a number of the more successful diagnostics that have been proposed to test for complement-hood/obligatoriness.2 Perhaps the best-known diagnostic is the elimination test (Helbig and Schenkel 1973)3 , in which the element in question is eliminated from the original sentence, and the derived sentence tested for grammaticality. In the case that a grammatical sentence is produced, the eliminated element is judged to have been non-obligatory, while if an ungrammatical sentence is produced, the element is seen to be syntactically obligatory. (3) a. b. c. The dog ate his dinner eagerly. * Ate his dinner eagerly. The dog ate eagerly.

Based on the elimination test, the dog is obligatory in (3) (and therefore a complement), but his dinner is optional. At the same time, no indication is given as to the complement/adjunct status of his dinner, pointing to the need for orthogonal diagnostics to fully characterise a given sentence element. In the case of Japanese, the applicability of the elimination test is highly restricted, in that most “obligatory” case slots are expressible with zero pronouns. Simple elimination of elements hence produces interpretational ambiguity as to the genuine non-existence of that element, and a zero anaphoric status, although disambiguation independent of sentential context is generally possible. While recognising this danger, however, we suggest that elimination at least provides an indicator in the obligatoriness evaluation process, and is valuable in comparing the eliminated contexts of di?erent arguments, as arguments higher in the argument status hierarchy (see below) produce more distinct markedness. A diagnostic which draws on the concept of “argument linking” is repeatability, suggested by Smith (1996:66). Here, Smith proposes that multiple instances of a given adjunct type generally produce grammaticality, whereas even if grammaticality is produced for multiple complement constituents, the produced interpretation will allocate distinct complement roles to each individual element. This
In this, we wish to distance ourselves from Uszkoreit’s (1987) claim of a full four-way adjunct/complement, optional/obligatory contrast in German. In terms of the valency binding hierarchy introduced below, Uszkoreit’s “obligatory adjuncts” are treated as middles. 2 Prominent tests not mentioned here include ‘back-formation’ (Helbig and Schenkel 1973), ‘substitution’ (Brinker 1972), ‘passivisability’ (Emons 1974), ‘addability’ (Heinz 1978) and the ‘do-so’ test (Somers 1984; Somers 1987). 3 Refer to Somers (1987:13-4) for a description of problems associated with the elimination test and the closely related test for ‘extractability’ (Brinker 1972).
1

3.1. COMPLEMENT/ADJUNCT DISTINCTION process can be exempli?ed with the following sentences. (4) *Tar? o-ga Hanako-ni Jir? o-ni atta Taro-nom Hanako-dat Jiro-dat met “Taro met up with Hanako and Jiro” (intended) Tar? o-ga Hanako-ni-mo Jir? o-ni-mo Taro-nom Hanako-dat-also Jiro-dat-also “Taro met up with both Hanako and Jiro” Tar? o-ga Hanako-to-Jir? o-ni atta Taro-nom Hanako and Jiro-dat met “Taro met up with Hanako and Jiro” tugi-wa 18-niti-ni 2-zi-ni kite-kudasai next-top 18th-dat 2 o’clock-dat please come “Please come next on the 18th at 2 o’clock” atta met

17

(5)

(6)

(7)

Naturally, this does not extend to cases of coordination for a given case position, but rather refers to separate surface realisations of that case slot. Hence, the acceptability of the in-case slot coordination in (6) and iterative -mo -mo construction in (5) do not threaten the integrity of the test. The process of a distinct complement role being forced on a repeated element can be seen with the verb pass below, whereby the salt in (8c) is forced into an Indirect Object role, producing a structure paralleling that in (8b). (8) a. b. c. Mary passed the salt Mary passed Peter the salt ?Mary passed the salt the pepper

One limitation of this diagnostic is its inability to account for the multiple-subject construction (Kuno 1973b:34, 68-78) in Japanese, where multiple nominative-marked constituents are generated in the Subject position. (9) ki-ga kirei desu yama-ga mountain-nom tree-nom pretty are ‘The mountains have beautiful trees’

adapted from Kuno (1973b:69)

Smith predicts this fact in identifying the nominative case marker as the default case marker in Japanese (p. 98). Because of this observation, however, repeatability is only applicable to non-subject case slots. One last test worthy of mention, which is speci?c to Japanese, is quanti?er ?oating (Kuroda 1980; Miyagawa 1988; Miyagawa 1989b). Quanti?er ?oating occurs when a numeric classi?er associated with a noun can be transposed to the right of the associated case slot. For this process to successfully occur, the noun phrase occupying the source case slot must necessarily be an obligatory element (Jacobsen 1992:41), and hence a complement. Quanti?er ?oating correctly identi?es the nominative case slot in the sentence pair of (10) as a complement. (10) gakusei-ga hutari student-nom two (people) “Two students came” kita came ?? hutari no gakusei-ga two (people) gen student-nom kita came

18

CHAPTER 3. VALENCY, ARGUMENT TYPES AND CASE

3.2

Argument status

This research hinges largely around the “argument status” of each case slot. In analysing argument status, we follow Bond and Shirai (1997) in combining the six-degree scale of valency binding proposed by Somers (1984, 1987), with the notion of ‘shadow complements’ as suggested by Pustejovsky (1995). The resultant argument status hierarchy is as follows. 1 2 3 4 5 6 7 Integral complement Shadow complement Obligatory complement Optional complement Middle Adjunct Extra-peripheral

Category 1 – Integral complements
Integral complements are generally constrained as to word order, are obligatory, have a lexically ?xed head, and are not anaphorically realiseable. Additionally, Japanese integral complements display only minor in?ection-driven case marker variation, with the accusative case alternating with the nominative for potential modality. However, there is often limited scope for modi?cation within the complement, multiplicity of case marking in the unmarked form is observable, and adverbial insertion can occur between the complement and verbal stem in restricted contexts. An example of an integral complement is the bucket in kick the bucket (taken from Bond and Shirai (1997)), or in the case of Japanese hidoime in hidoime-ni au “to have a hard time of it”. It is worthwhile noting that whereas the bucket cannot be modi?ed internally, hidoi-me can be modi?ed by the adverb monosugoku “extremely” to produce the accentuated idiom monosugoku-hidoime-ni au “to have a really hard time of it”, an e?ect which is equally applicable to the English gloss. Integral complements generate idiomatic verb sense, and multiple integral complement instances can occur within a single valency frame.

Category 2 – Shadow complements
Pustejovsky (1995:63-7) coined the term ‘shadow complement’ to describe ‘parameters which are semantically incorporated into the lexical item’ (p. 63), although the same phenomenon has been identi?ed independently by Talmy (1996:241) in his treatment of ‘blocked complements’. By this process, Pustejovsky refers to verbs such as butter, in which a default description of the material is entailed by the semantics of the verb, to the degree that a redundant surface realisation of that default material becomes ungrammatical (cf. (11b)). On the other hand, restricted instances of the default argument can be realised lexically (cf. (11c)), and replacement of the default argument by a synonym generally produces a grammatical usage (cf. (11d)). a. b. c. d. Mary buttered her toast. *Mary buttered her toast with butter. Mary buttered her toast with freshly churned butter. Mary buttered her toast with margarine. adapted from Pustejovsky (1995:65)

An example of a shadow complement from Japanese is zeikin-o “tax-acc” with the predicate n? ozeisuru “to pay tax”, which behaves analogously to butter above. A side-e?ect of classifying shadow complements lower in the hierarchy than integral complements is the misrepresentation of them being somehow less tightly bound to the predicate than integral

3.2. ARGUMENT STATUS

19

complements. In actual fact, as was seen for butter, they can be perceived as being so tightly bound to the predicate as to be non-realiseable in an unmarked form. However, unlike integral complements, synonym replacement is generally allowable (see (11d) above), and restrictions on word order are relatively relaxed in cases when a surface complement representation is possible. At the same time, we can observe some scope for in-complement modi?cation, as was observed for integral complements. Shadow complements are thus less rigidly restricted in surface form, supporting the given positioning below integral complements. An additional feature of shadow complements is the cline of acceptance of unmarked instances of the default complement in a matrix clause context. To take the above case from Japanese, zei-o n? ozei-suru was unanimously unacceptable to the native speakers consulted, whereas the matrix collocation of the shadow complement by? oin-ni “hospital-dat” with ny? uin-suru “to go into hospital/be hospitalised” received a relatively neutral response. Given that our usage of shadow complements is aimed at text analysis, we avoid making grammaticality judgements for such unmarked matrix occurrences of the default argument. Despite this relaxation of the constrained nature of shadow complements, however, we maintain a treatment independent of that for integral complements. That is, integral complements are necessary to derive the associated idiomatic sense of the predicate, in comparison with surface realisations of shadow complements which simply reinforce/extend the inherent verb sense generated by the predicate. This leaves open the question of the status of song in the construction sing a song. The default argument of song is unarguably encoded in the predicate, a fact which is evoked in the intransitive usage of sing, and only synonyms, hyponyms, and modi?ed instances of song are allowed as Direct Object. The exclusion of arguments in constructions of this type from the shadow complement classi?cation, stems from the acceptability of proper hyponym replacement, such as sing a shanty or sing a rollicking tune you heard on the radio. That is, semantic restriction on the Direct Object slot denotates a hierarchical semantic set of both synonyms and hyponyms, with the default of song at the root, unlike butter or zei “tax” which are highly restricted in themselves and are replaceable only with a limited range of synonyms, and modi?ed instances of that default sense. The inherent optional nature of shadow complements should be clear from the relative ungrammaticality of an unmarked surface occurrence of the default element. Application of the repeatability diagnosis, then, leads to the expected complement status of shadow complements, producing an optional complement (category 4) categorisation for shadow complements. On the surface, this would appear to cast doubt on the placement of shadow complements above obligatory complements within the valency binding hierarchy. We justify the given analysis from the observation that whereas elimination of surface shadow complements is possible, doing so reverts its semantic content to the default; for optional complements (see below), the same process of elimination simply leads to underspeci?cation. Moreover, any potential for synonym replacement is highly constrained, to a much higher degree than for obligatory complements. For the above reasons, the category 2 placement of shadow complements between integral complements and obligatory complements would appear to be well-founded.

Categories 3 & 4 – Obligatory/optional complements
Complements are de?ned as being ‘strictly subcategorised by the predicate for semantic class, syntactic function ... and often syntactic form’ (Somers 1987:28). Within this complement de?nition, however, we make a distinction between ‘obligatory’ and ‘optional’ complements, in line with the original valency-binding framework proposed by Somers (1987:27). In this, we diverge from the treatment given in (Bond and Shirai 1997). It is perhaps easiest to describe the di?erence between obligatory and optional complements by way of an example, in which the two complement types are indicated in bold and underlined, respectively.

20 (12) a. b. c. d.

CHAPTER 3. VALENCY, ARGUMENT TYPES AND CASE Peter handed over the documents. *Handed over the documents. *Peter handed over. Peter handed over the documents to the man in dark glasses.

Here, the elimination test clearly identi?es both Peter and the documents as obligatory complements, on account of the ungrammaticality of (12b) and (12c), respectively. In direct comparison, the man in dark glasses is an optional constituent. This third, optional constituent would, however, appear to play an equally salient role in the activity portrayed, in that it states the Recipient of the goods, in direct opposition to the Source of Peter. Subsequent assessment of the Recipient slot with the repeatability diagnostic described in section 3.1 clearly supports this intuition as to the complement status of the man in dark glasses. Through the combination of these above results, to the main in dark glasses can be seen to be an optional complement. Unlike the more restrictive complement types given above, however, ‘basic’ complements (i.e. complements not subsumed within the specialised categories 0, 1 and 2 given above) are compatible with anaphoric processes. In the case of Japanese, this equates to basic complements being realisable with zero pronouns, an e?ect which clouds the obligatory/optional complement dichotomy. Note that for English, the grammar requires that anaphoric derivatives of obligatory complements have a surface lexical form, a fact which can be used to determine obligatoriness. In fact, the obligatory/optional complement distinction in English basically corresponds to whether that complement is bound by the grammar to have a surface form, noting the deceitful behaviour of verbs which display distinct transitive and intransitive usages such as sing (see above): (13) a. b. Bill sang the national anthem with great gusto. Bill sang with great gusto.

The only tool we o?er to diagnose obligatoriness for a given complement position in Japanese is quanti?er ?oating (see above). The complement status of optional complements can then be veri?ed by way of repeatability.

Category 5 – Middles
Middles are proposed by Somers as an idiosyncratic ‘in-between’ classi?cation made up of elements which share the characteristics of both complements and adjuncts. Naturally they are non-obligatory, but the same close association can be observed with the governing verb. Examples of middles taken from English are Instrumental, such as with a hammer in hit the nail with a hammer, and Bene?ciaries, such as the squire in ‘The gamekeeper shot the squire a rabbit.’ (taken from Somers (1987:25)). These two case types also generally produce middle elements in Japanese. A signi?cant class of middles particular to Japanese is that of onomatopaeic adverbials, and notably phonomimes and phenomimes (Shibatani 1990:153-7). The strong correspondence between onomatopaeic expressions and particular verbs supports this view, as is seen for kusukusu “titter”, nikoniko “grin” and kutukutu “chuckle”, which collocate only with the verb wara(-u) “to smile/laugh”.4

Category 6 – Adjuncts
Adjuncts, again, are necessarily optional, but unlike middles tend to display semantic consistency across usage with distinct predicate classes. Naturally, pragmatic restrictions will exist as to local semantic compatibility with a given predicate, but, in general, their use is unpredictable. Co-occurrence of adjuncts of the same semantic type commonly occurs, as suggested by the repeatability test described above, and word order restrictions are relatively relaxed.
4

All these expressions also occur with the light verb suru “to do” in an inherent ‘laughing’ sense.

3.3. CASE SET

21

An example of an adjunct is the Japanese Locative case slot, such as kono-ie-de “this house-loc” in: (14) Tar? o-ha kono-ie-de umaresodatta Taro-top this house-loc was born and brought up “Taro was born and brought up in this house.” asu-no-hiru-ni kono-ie-de matiawasey? o tomorrow-gen-midday-dat this house-loc let’s meet “Let’s meet at this house at midday tomorrow.” kono-ie-de-wa uti-no-inu-ga itumo tukue-ya-isu-no-sita-de this house-loc-top us gen dog-nom always table and chair gen under-loc netagaru wants to lie “(When) in this house, our dog is always wanting to lie under tables and chairs”

(15)

(16)

Sentence (16) displays the repeatable nature of adjuncts.

Category 7 – Extra-peripherals
Extra peripherals are optional sentence modifying constituents, and constitute the outermost argument category; as one would expect, they are almost impervious to both word order and semantic restrictions. In both English and Japanese, adverbs form the main component of extra-peripherals. Particular examples are suddenly and often, and wazato “intentionally” and sorosoro “soon”.

3.3

Case set

To talk of case-role gapping, we must have a well-de?ned toolset of case-roles with which to tag case slots within the relative clause, and identify the source case slot in cases of gapping. The particular case-role paradigm employed in this research draws primarily on grammatical relations rather than strict case relations (cf. Fillmore’s (1968) Case Grammar, Starosta’s (1988) Lexicase, Somers’s (1987) case grid). By this is meant that the ‘core’ case-roles are syntactically relational, drawing on the notions of ‘Subject’, ‘Direct Object’ and ‘Indirect Object’. At the same time, however, the argument status hierarchy proposed in Section 3.1 is evoked in deriving the intermediate ‘Co-actor’ and ‘Co-patient’ roles from the traditional Direct and Indirect Object case-roles. In addition, the ‘grammatical’ basis applies only to the core case-roles, and the remaining eleven peripheral case-roles are de?ned based on semantic criteria. In order to validify the proposed case-role set, we endeavour to assign diagnostics to each case-role, as described in each case-role description below, although the reader is cautioned that these are speci?c to Japanese and no cross-lingual universality for either the case-roles or diagnostics is claimed. The proposed 18-way case-role set is grouped according to argument status and semantic similarity, into the core, other complement, local, temporal and oblique sets, as follows.5

3.3.1

Core grammatical cases

‘Core grammatical cases’ are complements (obligatory, optional or otherwise), and are de?ned surface syntactically, in the manner of Relational Grammar (Perlmutter 1980; Blake 1990).
For reference purposes, Machine Translation System Laboratory (MTSL) (1995), Nomura and Muraki (1996) and Bond and Shirai (1997) describe case-role sets for use in Japanese-English machine translation systems.
5

22

CHAPTER 3. VALENCY, ARGUMENT TYPES AND CASE

Subject
The Subject is traditionally de?ned as the general ‘doer’ of the action, such as the dog in The dog gnawed the rope. While this description is relatively uncontroversial for active clauses, it leads to two distinct treatments of passive subjects. The ?rst is to take a Fillmore Case-style approach and identify that entity which corresponds to the active subject, in the “underlying” or “logical” subject sense. The alternative method is to concentrate solely on surface syntactic marking in identifying the grammatical subject. Thus, in The rope was gnawed by the dog, the dog comprises the logical subject (coinciding with the grammatical subject of the active voice equivalent), and the rope the grammatical subject. In this research, we adopt this second, grammatical notion of subjecthood. Subjects are necessarily obligatory or integral complements, and are characterised by nominative case marking. Japanese Subjects cannot be unambiguously detected through either word order, in?ection, case marking, or the concept of surface syntactic obligatoriness. Rather, we must fall back on a number of linguistic tests to ascertain the Subject argument in a given sentence context. The ?rst such test is zibun-binding, and the observation that instances of the re?exive pronoun zibun can generally only bind to a clausal Subject position.6 (17) Tar? o-ga Hanako-ni Zir? o-o zibun no ie-de sy? okaisita. Taro-nom Hanako-dat Ziro-acc self gen house-loc introduced (lit.) “Taroi introduced Jiro to Hanako in self’si house.” (Shibatani 1990:283)

Hence, in (17), Taro is the Subject. One further test is subject honori?cation (Shibatani 1990:283), and involves the use of honori?c o V-ni naru marking on the main verb to indicate deference to the Subject: (18) a. syaty? o-ga waratta. president-nom laughed “The company president laughed” syaty? o-ga o-warai-ni natta. (Subject honori?c form of a.)

b.

Naturally, the Subject entity must be animate and pragmatically worthy of honori?cation for this test to be applicable. Despite this seemingly overbearing constraint, subject honori?cation provides an unambiguous means of determining the Subject case position through analysis of suitable situational participants, assuming that it is possible to uniquely identify one of those candidate participants as being worthy of honori?cation. That is, by identifying a case slot as containing a Subject ?ller in a given lexical context, for a given case frame, we can generalise that case slot as being the Subject in other lexical contexts, assuming consistency of case marking. An application of this process is the identi?cation of the Subject position of wakar(-u) “to know/understand” in a dative-nominative “ergative” marking context.7 First, it is necessary to generate a sentence context involving animate participants in all candidate case slots, one of which must be unambiguously superior in social standing to the others. Such a sentence is given in (19a), where syaty? o is the entity worthy of honori?cation. Next, we consider the “appropriateness” of subject honori?cation (cf. (19b)) and object honori?cation (cf. (19c)), and correlate these ?ndings with our a priori honori?cation judgement. In (19), the appropriateness of (19b) suggests the datively marked syaty? o as occupying the Subject position, leaving the nominatively marked syain as the Direct Object.
See (Iida 1996) for documentation of signi?cant exceptional cases in which zibun binds to non-subjects. Here again, note the non-coincidence between prototypical case-roles from the case marking types and the actual case-roles, with the Subject occupying a datively marked case slot and the Direct Object occupying a nominatively marked slot.
7 6

3.3. CASE SET (19) syaty? o-ni syain no koto-ga yoku wakaru. president-dat employee gen -nml-nom well understands “The president understands well his employees.” b. syaty? o-ni syain no koto-ga yoku o-wakari-ni naru. (Subject honori?c form of a.) c. * syaty? o-ni syain no koto-ga yoku zonziru. (Object honori?c form of a.)8 a.

23

Direct Object
Direct Objects generally indicate the entity/entities a?ected by the action described by the main verb. As such, they are expressible only as obligatory complements, or in terms of our argument status hierarchy, as obligatory or integral complements. One language-inspeci?c test for Direct Objects is that, in the absence of a Causee argument, they are commonly passivisable to the Subject position. This was the process observed above for the rope in The rope was gnawed by the dog. Direct Objects are prototypically marked with accusative case.

Indirect Object
Indirect Objects represent the ‘recipient’ or ‘bene?ciary’ in an action. Similarly to Direct Objects, Japanese Indirect Objects are passivisable (cf. (20), in which the Indirect Object is transformed into the Subject position), but only for (di)transitive verb senses (cf. (21)). (20) a. Tar? o-ga Hanako-ni tegami-o Taro-nom Hanako-dat letter-acc “Taro sent Hanako a letter.” Hanako-ga Tar? o-ni tegami-o Hanako-nom Taro-dat letter-acc “Hanako was sent a letter by Taro.” okutta. sent okurareta. was sent

b.

(21)

a.

b.

tegami-ga Hanako-ni watatta letter-nom Hanako-dat reached “The letter reached Hanako” *Hanako-ga (tegami-niyotte) watarareta Hanako-nom letter-by was reached “Hanako was reached by the letter.” (intended)

The default case marking for Japanese Indirect Objects is dative, and all Indirect Objects are optional complements.

Co-actors
In terms of traditional grammatical analysis, the Co-actor case-role straddles the boundary between the Direct and Indirect Object positions. It resembles the Direct Object case-role in argument status, in that all Co-actors are obligatory complements (but never integral complements). From the perspective of case marking, however, the Co-actor case slot is dative or comitative case marked, and hence most similar to Direct Objects. As an additional concern, Co-actors are not passivisable (cf. (22b)), in which respect they parallel intransitive Indirect Object usages.
8

Zonji(-ru) is a lexical object honori?c equivalent of wakar(-u).

24 (22) a.

CHAPTER 3. VALENCY, ARGUMENT TYPES AND CASE Tar? o-ga Hanako-ni/to Taro-nom Hanako-dat/com “Taro met Hanako.” atta. met

b.

*Hanako-ga (Tar? o-niyotte) awareta. Hanako-nom Taro by was met “Hanako was met by Taro.” (intended)

Under causativisation, on the under hand, Co-actors are transformed into Co-patients, (and hence coordinated with the Causee case-slot – see below): (23) Jir? o-ga Tar? o-o Hanako-to awa-se-ta. Jiro-nom Taro-acc Hanako-com meet-cause-past “Jiro introduced Taro to Hanako.” (lit. “Jiro caused Taro to meet Hanako.”)

One phenomenon which sets Co-actors apart from all Direct and Indirect Object usages is that Coactor case ?llers can be coordinated within the Subject case slot to retain the same basic sentential semantic (ignoring focus/theme variation). That is, they occur with reciprocal verbs and are mutually exchangeable. (24) Hanako to Tar? o-ga atta. Hanako and Taro-nom met “Hanako and Taro met.”

This con?ation of agency in the Co-actor case-role points to strong semantic resemblance between coordinating and comitative case marking roles of the particle ‘to’, although no claim is made as to the exact nature of this correspondence. The crucial di?erence between the two case-roles comes in the optional complement status of the Comitative, and it hence not being intrinsically de?ned for any verb. For example, while it is perfectly natural to say Tar? o-to(-issyo-ni) itta “(I) went along with Taro”, one would certainly not want to make the claim that the Comitative is de?ned within the valency frame for ik(-u) “to go”. This informal observation leads to the Comitative and Co-actor case-roles existing in disjunctive distribution. The optional nature of the Comitative case-role correlates to it being displaceable only with a dangling expanded comitative case marker in the relative clause body. (25) a. Jir? o-ga Tar? o-to(issyon-ni) p? ati-ni Jiro-nom Taro-com party-dat “Jiro went along to the party with Taro.” [ itta went

b.

Jir? o-ga issyo-ni p? ati-ni itta ] Tar? o Jiro-nom com party-dat went Taro (lit.) “Taro, who Jiro went along to the party with.” Jir? o-ga p? ati-ni itta ] Tar? o Jiro-nom party-dat went Taro “Taro, who Jiro went along to the party with.” (intended) *[

c.

3.3.2

Other complement cases

All ‘other complement cases’ are optional complements, with the sole exception of the Passive Agent, which is an obligatory complement for adversative passives.

3.3. CASE SET

25

Passive agent
The Passive Agent case-role is derived through passivisation, and marked datively or with ni-yotte.9 In the case of a “direct” passive (Miyagawa 1989b; Miyagawa 1989a), the Passive Agent is an optional complement, whereas Passive Agents in “adversative” passive contexts are obligatory complements (Hoshi 1994).

Causee
The Causee case-role is generated through causativisation, from Subject case slot transformation. It constitutes an obligatory complement (cf. Passive Agents) and is marked either datively or accusatively. (26) a. Tar? o-ga Hanako-ni hon-o yonda. Taro-nom Hanako-dat book-acc read “Taro read a book to Hanako.” Jir? o-ga Tar? o-ni Hanako-ni hon-o Jiro-nom Taro-dat Hanako-dat book-acc “Jiro made Taro read a book to Hanako.”

b.

yomaseta. made read

Co-patient
Co-patients mimic Co-actors in case marking (dative or comitative) and in their being non-passivisable (cf. (27b)), but di?er in that they coordinate with the Direct Object or Causee case-slots (rather than the Subject – cf. (27c)), are optional complements and are una?ected by causativisation. That is, Direct Objects being una?ected by causativisation leads to consistency of case-role coordination. (27) Tar? o-ga enzin-o haikikan-to kumiawaseta Taro-nom engine-acc exhaust pipe-com joined “Taro joined the engine with the exhaust pipe.” b. * haikikan-ga (Tar? o-niyotte) enzin-o kumiawasareta exhaust pipe-nom Taro-by engine-acc joined “The exhaust pipe was joined to the engine (by Taro).” (intended) c. Tar? o-ga haikikan to enzin-o kumiawaseta Taro-nom exhaust pipe and engine-acc joined “Taro joined the engine and exhaust pipe.” a.

Target
The Target case-role describes the optional target of experiential verbs (see Section 4.5.4) and is marked datively. (28) a. kabu no wariate-ga atta stock gen allotment-nom there was “There was an allotment of stock.” kozin-ni(taisite) kabu no wariate-ga atta individual-dat stock gen allotment-nom there was (lit.) “There was an allotment of stock to individuals.”

b.

The dative case marker and ni-yotte are essentially interreplaceable, excepting that ni-yotte cannot be used with adversative passives (Kuroda 1979). There is also a slight semantic di?erence between the two marking types, relating to the degree of “a?ectivity”of the Subject of the passive clause.

9

26

CHAPTER 3. VALENCY, ARGUMENT TYPES AND CASE

Object Allative
The Object Allative case-slot refers to the optional physical medium reference associated with resultative action verbs such as kak(-u) “to write (on)” and kake(-ru) “to hang (on)”; it is marked datively. (29) Tar? o-ga zibun no namae-o kami-ni Taro-nom self gen name-acc paper-dat “Taro wrote his name on the paper.” kaita wrote

3.3.3

Local cases

Japanese local case slots are consistently adjuncts. As suggested in the de?nition of adjuncts, this produces a relatively consistent spread of semantic variation across all patterns of usage, although predictable diversi?cation is seen in case marker collocation, according to the directionality of the local case slot. The relative consistency of semantic scope of local case usages means that a single “locative ?lter” can be utilised to analyse the locative compatibility of a given noun head/case ?ller. A default incompatibility with all local case types is assumed, unless otherwise marked; local case slots are generally una?ected by verbal modality, with the notable exception of the “resultative” -te ar(-u) auxiliary verb.

Locative
The Locative case slot construes the generic positional case-role, and is marked with the dative or locative (de) case markers. (30) Tar? o-ga T? oky? o-ni Taro-nom Tokyo-dat “Taro lives in Tokyo.” sundeiru. is living

(31)

kaigi-ga Meruborun-de okonowareta. conference-nom Melbourne-loc was held “The conference was held in Melbourne.”

Locatives can occur in either an ‘inner’ or ‘outer’ reading, with Inner Locative being the default. Unless otherwise stated, the term Locative will be used to refer to Inner Locatives throughout this paper. For a discussion of these Locative types and patterns of distribution, refer to Section 4.5.4.

Ablative
The Ablative case slot indicates the local source of a directional action, and is marked with the particle of the same name. (32) Hanako-ga harubaru Sapporo-kara yattekita. Hanako-nom all the way Sapporo-abl came “Hanako came all the way from Sapporo.”

Allative
The Allative case slot indicates the local target of a directional action, and is marked datively or by the allative e/made case particles. (33) Tar? o-ga sengetu Ky? oto-made/ni/e borosya-de unten-site kita. Taro-nom last month Kyoto-all-dat-all old car-loc drove there and back “Taro drove to Kyoto and back in his bomb (of a car) last month.”

3.3. CASE SET

27

Perlative
The Perlative case slot indicates the location through/across which the directional action of the main verb occurs, and is marked accusatively. (34) Tar? o-wa izen tomodati no hune-de Nihonkai-o Taro-top previously friend gen boat-loc Japan Sea-acc “Taro has previously crossed the Japan Sea in a friend’s boat.” watatta-koto-ga aru. has crossed

3.3.4

Time cases

Japanese time cases are used for time-related reference. They are expressed as adjuncts and marked datively or with null marking. By default, verbs are assumed to be compatible with time case reference.

Durational
The Durational case-role indicates the length of time an action or state lasted. Unmarked durational nouns such as kikan “interval” are generally datively marked (with or without a pre-dative zy? u durational marker), whereas durational complexes involving cardinal reference are most commonly associated with null marking. (35) s? eru no kikan-ni mise-o odozureta sale gen interval-dat shop-acc visited “( ) visited the shop during the sale.”

Temporal
The Temporal case-role indicates a distinct point in time, and is marked with the dative case marker for cardinal date/time references, and with null/iterative (mo) marking for generic temporal expressions such as ky? o “today” and kyonen “last year”. (36) ky? o Tar? o-ga kuru. today Taro-nom comes (lit.) “Today, Taro will come.” mikka-ni Tar? o-ga kuru. 3rd-dat Taro-nom comes (lit.) “On the third, Taro will come.”

(37)

3.3.5

Oblique cases

Instrumental
The Instrumental case-role is used to describe a tool or instrument used in performing an action. The extension of tool types is highly verb-speci?c, supporting a middle status and default incompatibility with Instrumental reference. Instrumentals are marked with the locative (de) case marker. (38) Tar? o-ga hasami-de tegami-o aketa Taro-nom scissors-loc letter-acc opened “Taro opened the letter with a pair of scissors.”

Cardinal
The quantity/degree of an action is expressed with the Cardinal case-slot, with the semantic scope spanning from unit-based mention such as sokudo “speed”, to physical extent and price, such as with

28

CHAPTER 3. VALENCY, ARGUMENT TYPES AND CASE

kakaku “price”. Cardinals are marked with the locative case marker or null marking, are adjuncts, and are by default incompatible??? with all verbs. (39) hik? oki-ga monosugoi hayasa-de sora-o tonda plane-nom extreme speed-loc sky-acc ?ew “The plane ?ew across the sky at extreme speed.”

Chapter 4

Verb class-based resolution
4.1 Verb class hierarchy

The most fundamental mechanism called upon in realising this research is the derivation of a verb class hierarchy to cluster verbs based on valency, argument preferences, inter-case-role relations and argument type preferences. By way of linking verb classes to valency variation and inter-case-role relations, it is possible to slot optional arguments in and out of the valency frame as required, and apply inter-case-role preferences. The way this ‘linking’ is achieved in this research is that minirule sets are stipulated a priori for each verb class, and applied in parallel for the full verb class characterisation of the verb in question to produce multiple analysis types. Such rule sets are made up of if-then-else conditionals, frequently consisting of a single rule. Each rule set is designed to produce at most one output, with semantic incompatibility potentially leading to failure of any rule being triggered. The various outputs are then combined based on analytical coincidence, and weighted variously to produce a unique relative clause interpretation (see Chapter 6).

4.1.1

Argument status and interpretation preference

Commonly, the use of valency frames in a task of this type is closely linked to the introduction of sortal preferences for each case slot modelled, so as to be able to di?erentiate between distinct senses for a given verb stem (verb sense disambiguation ), and at the same time disambiguate the case-role of each argument (case-role disambiguation ). In the case of our system, however, the decision was made to alleviate local sortal preferences, so as to avoid consideration of verbal polysemy as much as possible. In return, argument status is used as an indicator of interpretive salience/accessibility, such that lower argument status interpretations will generally override higher value interpretations. For example, given an integral complement interpretation and an adjunct type interpretation, the integral complement will be preferred outright. The top level argument preference set is thus: Full clause-based idiom argument types Integral complement Shadow complement Middle Other

That is, full-clause based idioms are strictly preferred over integral complements, which in turn take precedence over shadow complements, and so on. The placement of middles above obligatory and optional complements may seem controversial in light of the higher argument status of these types. However, the well-de?ned lexical nature of middles makes them more stable through relativisation than general complements. The remainder of argument types (obligatory/optional complements, and adjuncts) interplay on a ?ner level, with selection between obligatory and optional complements being made from the mappings 29

30

GENERIC

Full clause-based idiom

Quotative EVENT

Quantative

VALENCE VARIATIONAL

STATE Start End ACTION

Conflated ergative

Partitive

Copula Natural PHYSICAL Extinction/ phenomenon MOVEMENT destruction

Mental Tool-aided Locational action action action RELATIONAL

Conjoining

Existential

Experiential

Distal movement Proximal Travelling movement

Generic relational

Inter-personal relational

PART-WHOLE RELATIONAL

CHAPTER 4. VERB CLASS-BASED RESOLUTION

Figure 4.1: The full verb class hierarchy (Original verb classes indicated in bold, partitioning nodes capitalised)
Empathy Including Excluding

4.1. VERB CLASS HIERARCHY

31

between them inherent in the valency frame, and adjuncts weighted according to the prototypicality of the noun head with that adjunct type. We claim that the proposed treatment of adjunct case slots is one of the potential strengths of the system, in terms of consistency of application and simplicity of processing. Given that adjuncts behave relatively consistently across all usages, it is possible to simply de?ne adjunct type compatibilities for each adjunct case-role, and apply a uniform semantic treatment to the calculation of adjunct correspondence of arguments. In terms of verb class representation, this equates to associating verb classes to each adjunct type, determining a default (in)compatibility judgement (see Section 3.3), and marking those cases for which the particular verb sense does not coincide with this judgement.

4.1.2

Constructing the verb class hierarchy

In constructing the verb class hierarchy, the NTT “verbal semantic attribute” (‘verb class’ hereafter) hierarchy (Nakaiwa et al. 1994; Nakaiwa and Ikehara 1997) provided a solid starting point, which we were able to expand upon and modify for our purposes. Within the NTT transfer dictionary, verb classes are included as a general representation of the semantic type of each verb sense, in a cognitive or real world-applicability sense. They are called upon in the ALT J/E system during the discourse processing of zero pronouns (Nakaiwa and Ikehara 1994; Nakaiwa and Ikehara 1995; Nakaiwa et al. 1995; Nakaiwa and Ikehara 1996), and in English article and classi?er generation. However, due to the principal intention of the verb class hierarchy to document clause-level semantics, rather than intra-clausal case-role relations, the verb class hierarchy does not in itself meet the lower level representational needs of our system, in replacing the need for case slot-based type constraints. Ultimately, therefore, existing verb class descriptions were combined in the single default valency frame entry generated for each verb stem (see Section A.1.1), and complemented with additional verb class types devised independently for our relative clause analysis purposes.

Verb class multiplicity
Multiplicity of verb class characterisation exists to a limited extent in the original NTT verb entries (Nakaiwa and Ikehara 1997:219), with the original average number of verb classes associated with each entry at 1.12. For the dictionary used in our system, the average number of verb classes per entry rose signi?cantly to 1.45, with the distribution of verb class frequencies across the dictionary as given in Table 4.1. This multiplicity of rule class categorisation complicates the relative clause analysis process somewhat, as we can often expect multiple verb class-based analyses for the relative clause, leading to scope for multiple analyses. Weighting and ranking schemes interfaced to the various rule sets are described in Chapter 6, but su?ce to say that all analyses are combined additively, and the degree of semantic compatibility is weighted appropriately. The ?nal output is then the analysis type which receives the greatest overall score.

Our verb class hierarchy
The ?nal verb class hierarchy is as presented in Figure 4.11 , in which labels indicated in bold are verb classes developed independently in this research, and NTT verb classes to which a particular rule set has been attached; capitalised labels are dummy nodes used to delineate/structure the verb class hierarchy but which have no verb class semantic in themselves. The hierarchy presented does not detail the full extent of the 107 verb classes contained within the full NTT transfer dictionary and maintained in the verb entry extraction process, but rather those classes/class clusters speci?cally
1

Many thanks to Hiromi Nakaiwa of NTT for providing English translations of the full range of NTT verb classes.

32 No. verb classes for entry (Total: 11533) 1 2 3 4 5 6 7 8 9 10 14

CHAPTER 4. VERB CLASS-BASED RESOLUTION Frequency (Total entries: 7951) 5463 1798 457 144 52 16 11 4 2 3 1 Table 4.1: Distribution of verb class frequencies Proportion

68.71% 22.61% 5.75% 1.81% 0.65% 0.20% 0.14% 0.05% 0.03% 0.04% 0.01%

utilised in this research. For details of the “full clause-based idiom” class, the reader is referred to Section 2.3.3. Individual descriptions of the newly developed verb classes, and associated rule sets where applicable, are provided below.

4.2

Physical movement

As the name suggests, physical movement verbs indicate a directional movement, as distinct from stasis. This can take the form of distal movement towards a particular location, proximal movement away from a particular location, or travelling motion through a locus.

4.2.1

Distal movement

Distal movement verbs are associated with compatibility with the Allative case-role, realised interchangeably through the dative and allative (e/made) case markers. Examples of distal movement verbs are ik(-u) “to go” and muka(-u) “to head towards”. IF (locative head AND uninstantiated Allative case slot) RETURN Allative;

4.2.2

Proximal movement

Proximal movement verbs are associated with compatibility with both the Ablative and Allative case-roles. Case marking for the Ablative case-role involves either the accusative or ablative (kara) case marker, and Allative case-role marking is identical to that for distal movement verbs, that is dative or allative. The default local case-role, however, is the Ablative, and Allative case-roles are generally only tenable with lexically marked allative nouns such as saki “destination”. That is, the focus of proximal movement verbs is on the starting point of the action, and any unmarked locative is automatically associated with the Ablative case-role, but this preference can be overridden given overt allative-type marking. Examples of proximal movement verbs are ririku(-suru) “to take o?” and t? ozakar(-u) “to recede/go away (from)”.

4.2. PHYSICAL MOVEMENT

33

Verb class

Frequency (Total entries: 7951) 79 25 3 124 6 76 1 126 3 77 289 18 3 689 1309 2173 101 59 49 13 160 72 22 66

Percentage of entries containing given verb class 0.99 0.31 0.04 1.56 0.08 0.96 0.01 1.58 0.04 0.97 3.63 0.23 0.04 8.67 16.46 27.33 1.27 0.74 0.62 0.16 2.01 0.91 0.28 0.83

Con?ated ergative Conjoining Copula Distal movement Empathy End Excluding Existential Experiential Extinction/destruction Generic relational Idiom Including Inter-personal relational Locational action Mental action Natural phenomenon Quantative Partitive Proximal movement Quotative Start Tool-aided action Travelling

Table 4.2: Verb class frequency

34

CHAPTER 4. VERB CLASS-BASED RESOLUTION IF (allative noun head) RETURN Allative; ELSE IF (locative head AND uninstantiated Ablative case slot) RETURN Ablative;

4.2.3

Travelling

The focus for travelling verbs is on the Perlative case-role, and the route taken in the travelling motion; this is marked with the accusative. However, similarly to proximal travelling verbs, there is potential to refer to the destination through the use of ablative nouns, for which the case marking is allative or dative. T? or(-u) “to travel/pass through” and tob(-u) “to ?y (across)” are both instances of travelling verbs. IF (allative noun head) RETURN Allative; ELSE IF (locative head AND uninstantiated Perlative case slot) RETURN Perlative;

4.3

Relational

Relational verbs are characterised by relating a source and target entity. In the case of inter-personal relational verbs, both entities are generally human, whereas generic relational verbs are associated with a broader range of both animate and abstract arguments. For all relational verbs, the focus is on the source entity, which in the context of relative clauses means that the default gapped caserole in cases of ambiguity between the source and target case slots, corresponds to the source. For target case-role gapping to occur, one or more of the following conditions must be met: (a) the source entity must be lexically realised, (b) the head must be an allative noun, or (c) there must be marked empathy on the source entity (see below). The most commonly occurring personal allative noun is aite “opponent”, although the non-personal allative saki “direction/goal” can be equally acceptable for generic relational verbs in certain contexts. Target entities can occur in any of the Co-actor, Co-patient and Indirect Object case-roles, with the particular case-role de?ned by the predicate and exclusivity of these case-roles occurring in a given valency frame. Co-actor targets are obligatory in nature (a fact which derives directly from the de?nition of the Co-actor case-role), which makes them slightly more tenable to unmarked case-role gapping than the other two target argument types. They occur for “reciprocal” verbs such as a(-u) “to meet” (interpersonal relational) and itti(-suru) “to correspond” (generic relational). The ‘reciprocity’ of Co-actor target elements can be observed in (1), where ambiguity exists between Subject and Indirect Object case-role gapping. (1) au ] hito meets person a. “people who meet ( )” b. “people ( ) meets” [

Clearly, the two glosses correspond to the same situation, and if there is to be any constraint on the two case-roles, it is that the most topical/empathised entity occupies the Subject case slot for interpersonal relational verbs. We return to this matter in discussion of the system evaluation in Section 8.2.4.

4.3. RELATIONAL

35

Similarly to Co-actors, Co-patients are obligatory and occur for reciprocal-sense verbs. Here, however, the relational correspondence occurs with the Direct Object case slot, with di?ering degrees of reciprocity. In the case of kumiawase(-ru) “combine”, for example, full interreplaceability is possible, whereas complications occur for ‘replacement’-sense verbs such as tyenzi(-suru) “to change over” and k? okan(-suru) “to replace”. Even with k? okan(-suru), however, a higher degree of reciprocity is seen than for the English replace, in that the Direct Object source element can indicate the replacing item given appropriate replacing-type markedness on the case ?ller.2 Thus, while recognising that implicit directionality is evident for certain Co-patient marked source types, the Direct Object and Co-patient case slots can generally be interchanged. Indirect Object-type targets generally refer to the Recipient or Bene?ciary of the described action, and are optional (again, obtained from the de?nition of Indirect Objects). Unlike Co-actors and Co-patients, Indirect Object targets produce a de?nite sense of directionality of the action, are not reciprocal (seen in the non-equivalence of (2a) and (2b)), and cannot be coordinated with the target case-role while retaining the same sense, as occurred above for Co-actors and Co-patients (cf. (2c)). Examples of Indirect Object targets occur with the verbs watas(-u) “to hand over” and aisatu(-suru) “to greet”. (2) a. Tar? o-ga Hanako-ni tegami-o okutta. Taro-nom Hanako-dat letter-acc sent “Taro sent Hanako a letter.” b. Hanako-ga Tar? o-ni tegami-o okutta. Hanako-nom Taro-dat letter-acc sent “Hanako sent Taro a letter.” (= a.) c. Hanako to Tar? o-ga tegami-o okutta. Hanako and Taro-nom letter-acc sent “Hanako and Taro sent a letter (to ( ) ).” (= a.)

4.3.1

Inter-personal relational

Inter-personal relational verbs relate two animate entities, in the source case slot α (the Subject or Direct Object) and target case slot β (any of the Co-actor, Indirect Object and Co-patient case slots). The only personal allative considered is aite. For details of the determination of the animacy of the noun head, see Section 6.3. IF (animate head) IF (personal allative noun head AND uninstantiated target case slot β ) RETURN β ; ELSE IF (uninstantiated source case slot α) RETURN α; ELSE IF (uninstantiated target case slot β ) RETURN β ;

Empathy
‘Empathy’ verbs (Kuno 1978) form a proper subset of inter-personal relational verbs, and are de?ned by their incompatibility with a ?rst person pronoun in the target case slot for simple in?ection usages. (3) * Tar? o-ga watasi-to Taro-nom I-com “Taro met with me.” atta. met

Mechanical and civil engineering-related instruction manuals frequently contain sentences such as atarasii boruto-o k? okansuru (new bolt-accreplaces) “replace ( ) (with) the new bolt”, in syntactically unmarked usages.

2

36

CHAPTER 4. VERB CLASS-BASED RESOLUTION

This high degree of empathic focus on the source entity produces the e?ect that target case-role gapping can occur without a surface realised source entity, for unmarked head nouns. Unfortunately, this often leads to lack of focus-based preferences between the source and target case slots in the case that both are uninstantiated, unless the head noun intension is marked allatively or empathically. At the same time, however, temporal or locative grounding tends to weight the focus towards the target case slot, as does the past tense. (4) a. [ au ] hito meets person “the person who met ( nitiy? obi-ni Sunday-dat “the person ( [ [

) ” vs. “the person (

) met”

b.

atta ] hito met person ) met on Sunday”

c.

Nihon-de au ] hito Japan-loc meets person “people (one) meets in Japan”

The handling of this marginal preference for the target case slot is a somewhat brutal one, in that simple existence of past tense in?ection or local grounding is seen to generate unambiguous target gapping. However, given that there are no factors working to reverse the preference back in the other direction, the given treatment seems su?cient. The algorithm for empathy verbs interfaces with that for inter-personal relational verbs through sequentiality, in that the following rule is applied prior to the inter-personal relational verb algorithm, and if an output is returned, that analysis type is automatically returned from the inter-personal relational verb algorithm. IF (animate head AND uninstantiated target case slot α AND relative clause is in past tense or is locally grounded) RETURN α;

4.3.2

Generic relational

Generic relational verbs are identical to inter-personal relational verbs, except that there is no semantic restriction on the source and target case slots. The rule set is thus basically the same as that for interpersonal relational verbs (see above). IF (allative noun head AND uninstantiated target case slot β ) RETURN β ; ELSE IF (uninstantiated source case slot α) RETURN α; ELSE IF (uninstantiated target case slot β ) RETURN β ;

4.3.3

Part-whole relational

Part-whole relational verbs are subdivided into the including and excluding verb classes, and exemplify/exclude some part of the whole characterised by the modi?ed noun head.

4.3. RELATIONAL

37

Including
Verbs contained in the including verb class can be used in a non-restrictive exempli?cation form, realisable in the simple non-past or past tense. The exemplar set is construed in the accusative case3 , which must be present to trigger the including sense, and no further arguments can collocate with the main verb. ? u (5) [ t? ozai-o fukumeta ] Osy? east and west-acc included Europe (lit.) “Europe, including (both) the east and the west” Including relative clauses are head restrictive, and the system output on detection of this relative clause type is the tag for this modifying type, i.e. Inclusive. The fact that including relative clauses are not case-slot gapping can be seen by considering a simplex derivation of (6a) below. (6) a. [ Beisut? azu-o fukumu ] zen yaky? u t? imu Baystars-acc includes all baseball teams “all baseball teams, including the Baystars” zen yaky? u t? imu-ga Beisut? azu-o fukumu all baseball teams-nom Baystars-acc includes “All baseball teams include the Baystars.”

b.

Clearly, the scope of the existential quanti?er zen “all” is not equivalent between (6a) and (6b), excepting the case where the Subject in (6b) is treated as being textually containing through quotation, thus restricting the scope of the quanti?er to the Subject NP. However, this quotative interpretation is not available in (6a) and hence does not constitute direct equivalence. Members of the including verb class include fukume(-ru) and hazime-to(-suru). IF (simple main verb inflection AND unique accusatively marked argument) RETURN Inclusive;

Excluding
Excluding verbs extensionally restrict the modi?ed head noun by identifying elements which are to be excluded from the default denotation. The exclusion sense of these verbs is produced for simple tense usages with only the accusative case slot instantiated.4 (7) nitiy? o-o nozoku ] mainiti Sunday-acc excludes everyday “everyday, excluding Sundays” [

While usages such as (7) can be related back to the unmarked simplex sense nozoku, scope di?erences occur between excluding relative clauses and the corresponding simplex clause derivant, as was seen for the including verb class above: (8) mainiti-kara nitiy? o-o nozoku everyday-abl Sunday-acc excludes “to exclude Sunday from every day”

Verb arguments can also be marked with the iterative (mo) marker. Note that for excluding verbs, the unique verb argument cannot be marked with the iterative (mo) marker, unlike including verbs.
4

3

38

CHAPTER 4. VERB CLASS-BASED RESOLUTION

This supports a head restrictive relative clause treatment for excluding relative clauses. The excluding verb class consists uniquely of the verb nozoku. IF (simple main verb inflection AND unique accusatively marked argument) RETURN Exclusive;

4.4

Valence variational

Valence variational verbs are associated with distinct valency values, with a well-de?ned case slot mapping between the di?erent valency frame types. In traditional valency frame dictionary treatments, such correspondences are not represented explicitly, and separate entries are simply allocated for the di?erent valency types. The inadequacy of this method is described below for the various valence variational sub-types.

4.4.1

Con?ated ergative

Con?ated ergative verbs have been independently identi?ed by Jacobsen (1992:212) as a noteworthy exceptional case in his proposed transitivity account of Japanese. As identi?ed by Jacobsen, con?ated ergative verbs are unambiguously of the Sino-Japanese type and display both transitive and intransitive usages, in ergative correspondence. (9) a. Tar? o-ga sagy? o-o Taro-nom work-acc “Taro started work.” sagy? o-ga kaisi-sita work-nom started “Work started.” kaisi-sita started

b.

Considering this e?ect from the perspective of relative clauses, inherent ambiguity of valency frame type arises in cases where no overt Subject is given. (10) Tar? o-ga kaisy? o-sita ] mondai Taro-nom solved problem “the problem Taro solved” b. [ kaisy? o-sita ] mondai solved problem “the problem ( ) solved” vs. “the solved problem” c. [ kaisy? o-sita ] hito solved person “the person who solved ( )” a. [

In (10a) above, that the transitive sense is evoked is recoverable from the animate Subject, forcing mondai into the Direct Object position. For (10b), on the other hand, it is not possible to determine the transitivity of kaisy? o(-suru) through sortal preferences, as they apply equally to the Subject case slot for the intransitive sense and Direct Object case slot for the transitive sense (in which case a zero subject interpretation would be produced). In fact, the only means of resolving this con?ict is discourse processing, and the determination of a suitably salient discourse entity, capable of ?lling the transitive Subject position. Failure to locate such an entity would point to intransitive valency, and successful identi?cation of an animate Subject would suggest transitive valency. Given the current discourse-independent nature of our relative clause treatment, this type of high level processing cannot be called upon. The solution to the problem is thus to assume the unmarked

4.4. VALENCE VARIATIONAL

39

intransitive interpretation in the case that an overt Subject is not supplied within the relative clause. This correlates to hedging on the transitivity issue, as no assumption is made one way or the other as to the full content of the valency frame, and the resultant Subject analysis is equally applicable to both intransitive and transitive Subject analysis. Indeed, the only instance in which this analysis would prove incorrect is where gapping has occurred from the Direct Object case slot of the transitive sense of the verb in question, for a zero Subject. IF (uninstantiated Subject case slot) RETURN Subject; ELSE RETURN Direct Object;

4.4.2

Partitive

Partitive verbs contain a “part”/attribute in the nominatively marked Subject case slot, and can optionally collocate with a topic-marked “whole”, to which the clausal attribution applies. (11) a. iro-ga aseta. colour-nom faded “The colour faded.” s? et? a-wa iro-ga aseta. sweater-top colour-nom faded “The colour faded out of the sweater.”

b.

The whole and part can alternatively be coordinated in the Subject position, suggesting the clause initial topic construction as a “major subject” (Tateishi 1994), and hence a displaced ‘whole’ in a relative clause context as generating a bound gapping clause. (12) s? et? a no iro-ga aseta. sweater gen colour-nom faded (lit.) “The colour of the sweater faded.”

In terms of our case-role schema, the optional “whole” topic is classi?ed as a (second, anchoring) Subject. For relative clause analysis, the gapping of the anchoring “whole” case slot translates to a bound gapping relative clause instance, with Subject gapping. Gapping from the “part” case slot in the absence of the “whole”, on the other hand, constitutes simple Subject gapping, noting that gapping of the part in the presence of the whole is not possible. (13) [ iro-ga/no aseta ] s? et? a colour-nom/gen faded sweater “a sweater which has faded in colour”

(14)

aseta ] iro faded colour “a faded colour” s? et? a-ga aseta ] iro sweater-nom faded colour “the color which faded from the sweater” (intended) *[

[

(15)

The ungrammaticality of relative clause complexes of type (15), combined with the constraint that gapping of the part can occur only in the absence of the whole, means that (14) is not associated with the pragmatically acceptable second interpretation of (lit.) the colour which faded from ( ). However, analysis of (14) by means of the valency frames attributed to the simple ‘part’ and complex

40

CHAPTER 4. VERB CLASS-BASED RESOLUTION

‘whole-part’ senses would produce this second interpretation, of strictly equivalent acceptability to a faded colour due to the identical type restrictions on the ‘part’ case slot in the two frames. For this reason, the partitive verb class would appear semantically justi?ed. The current algorithm is limited in its potential to capture correspondences of this type, by the lack of a broad-coverage world knowledge source, with which to derive part-whole relationships. Thus, the actual handling of “whole” Bound Subject gapping of the type given above, is simpli?ed to assume that Bound Subject gapping occurs only in the context of full complement case instantiation. What this means in real terms is that given full complement case instantiation, the system should prefer a Bound Subject interpretation over a head restrictive relative clause interpretation, these two analysis types being the only two possible alternatives. Clearly, therefore, considerable scope exists to improve the current treatment of partitive verbs, and this is left as an item for future research. IF (fully instantiated complement case frame) RETURN Bound Subject;

4.5
4.5.1

Other verb classes
Copula

The copula verb class is used to tag the various realisations of the Japanese copula. The copula is relatively orthodox in behaviour within the context of relative clauses, in that case-role gapping can only occur from the Subject case slot, and incompatibilities exist with both temporal and local case-roles.5 The copula verb can take the forms de-ar(-u), de-gozar(-u), de-irrasya(-ru) and da. IF (Subject case slot uninstantiated) RETURN Subject;

4.5.2

Conjoining

Conjoining verbs closely resemble the copula from the standpoint of case-role gapping, by way of gapping only from the Subject position and being incompatible with both temporal and local caseroles. As implied by the nomenclature, conjoining verbs semantically ‘conjoin’ or ‘relate’ concept pairs, but di?er from relational verbs in that gapping cannot occur from the target (non-subject comparator) case slot. Examples of conjoining verbs are uwamawar(-u) “to exceed” and kanren(-suru) “to relate to” IF (Subject case slot uninstantiated) RETURN Subject;

4.5.3

Quantative

Quantative verbs are exempt from the default adjunct compatibility for time and cardinal adjuncts, with quantative arguments implicitly expressible through the valency frame-de?ned ‘maximally peripheral complement case slot’. In real terms, the maximally peripheral complement case slot is the ?nal (rightmost) complement represented within the valency frame.
Due to the modular nature of the given verb class system in handling adjuncts, it is neither possible nor desirable to code multiple adjunct (in)compatibilities within a single verb class, and the observed adjunct incompatibilities for copula verbs are not applied directly from the copula verb class within our system.
5

4.5. OTHER VERB CLASSES (16) [ kakatta ] zikan took time “the time taken”

41

For (16), the noun head of zikan “time” would, by default, point to a Temporal or Durational caserole gapping analysis. The quantative membership of kakar(-u) overrides this analysis type, however. Instead, the nominatively marked Subject case slot is identi?ed as the ‘quantative case slot’, and an attempt is made to map zikan onto the Subject case slot. Finding that the Subject case slot is uninstantiated, the system correctly returns a Subject analysis for the relative clause complex. Examples of quantative verbs include kakar(-u) “to cost/take (time)” (maximally peripheral/quantative argument case slot = Subject) and tuiyas(-u) “to consume” (maximally peripheral/quantative argument case slot = Direct Object) . IF (quantative noun head AND uninstantiated maximally peripheral complement case slot α) RETURN α;

4.5.4

Existential

Existential verbs are stative verbs which can include mention of the locus??? of the state. As a direct consequence of the adjunct status of the locative case-role, multiple mention of locus can occur, with the separate locative case slots marked distinctly as being ‘inner’ and ‘outer’ positions (inner/outer terminology taken from Halliday (1970) and Platt (1971)). The (basic) Inner Locative is marked datively and is the default, whereas the (peripheral) Outer Locative is marked with the locative case marker (de), and occurs only in conjunction with the Inner Locative. This marks a point of departure from prototypical adjunct repetition, by way of repeatability not extending to the case marking level. Allocation to the two case slots is determined according to the relative speci?city or local granularity of the locative case ?llers, with the ?ner grained case ?ller occupying the inner case slot. (17) Tar? o-ga Pari-de(ha) okina ikkenya-ni ? Taro-nom Paris-loc(top) large house-dat “Taro lives in a big house in Paris.” sundeiru is living

In (17), for example, ? okina ikkenya corresponds to the Inner Locative case slot, and Pari to the outer case slot. The role of granularity in demarking these case slots is evident in that ? okina ikkenya is geographically contained within the extension of Pari, and the two locatives can be coordinated by use of the genitive connective (no) producing Pari no ? okina ikkenya “big house in Paris”. Plugging this genitive coordinated locative back into the original clause, we see that the two locatives are con?ated within the Inner Locative case slot: (18) Tar? o-ga Pari no ? okina ikkenya-ni Taro-nom Paris gen large house-dat “Taro lives in a big house in Paris.” sundeiru is living

Additionally, if we consider (17) in the absence of the mention of ? okina ikkenya, we see that Pari is forced from the Outer Locative case slot into the Inner: (19) a. * Tar? o-ga Pari-de sundeiru Taro-nom Paris-loc is living “Taro lives in Paris.” (intended) Tar? o-ga Pari-ni sundeiru Taro-nom Paris-dat is living “Taro lives in Paris.”

b.

42

CHAPTER 4. VERB CLASS-BASED RESOLUTION

From this, it is clear that the outer locative case slot occurs only in simplex conjunction with an Inner Locative, and conversely that any singular locative case-role mention for existential verbs must occur in the Inner Locative case slot (and hence be datively marked for existential verbs). Returning to consideration of relative clauses, this produces an immediate result for locative gapping existential verbs. That is, the head of a relative clause containing an Outer Locative and no Inner Locative must have been gapped from the Inner Locative case slot, given that the Inner/Outer Locative dichotomy is preserved under case-role gapping. Hence, given that an Outer Locative can only exist in the presence of an Inner Locative mate, ‘dangling’ Outer Locatives indicate cases of Inner Locative gapping. (20) Tar? o-ga Pari-de sundeiru Taro-nom Paris-loc is living “the house Taro lives in in Paris” [ ] ie house

Additionally, in the absence of any locative in the relative clause body, Locative gapping must occur from the Inner Locative case slot, noting that the Outer Locative case slot is equally tenable to case-role gapping as the Inner Locative. (21) Tar? o-ga okina ikkenya-ni sundeiru ] ? Taro-nom big house-dat is living (lit.) “Paris, where Taro lives in a big house” [ Pari Paris

Examples of existential verbs are sum(-u) “to live/inhabit” and kizon(-suru) “to exist”. IF (locative head) IF (Inner Locative case slot uninstantiated) RETURN Inner Locative; ELSE IF (Outer Locative case uninstantiated) RETURN Outer Locative;

Experiential
Experiential verbs6 form a proper subset of existential verbs, and are additionally compatible with an optional Target case slot, realised in the dative case. Examples of experiential verbs are i(-ru) “to be/have” and ar(-u) “to be/have”. IF (uninstantiated Subject case slot) RETURN Subject; ELSE RETURN Target;

4.5.5

Locational action

Locational actions closely resemble experiential verbs, in that they are actions which can be attributed a locus or locii of action, although here, case marking is in the locative case (de) for both Inner and Outer Locatives. This leads to one major di?erence with existential verb-based relative clauses, in that a single Locative within the relative clause body is automatically interpreted as the Inner Locative. Thus, Inner Locative gapping only arises in the complete absence of any locative argument, and the Inner Locative gapping relative clause equivalent of a double Locative matrix clause produces ungrammaticality7 :
We refer to experiential verbs as ‘stative’ in (Baldwin et al. 1997b). Note that grammaticality is produced for the bound Locativerelative clause equivalent, due to the preservation of the Inner and Outer Locative roles:
7 6

4.5. OTHER VERB CLASSES (22) a. kono ie de-wa uti no inu-ga itumo tukue ya isu no sita-de this house-loc-top us gen dog-nom always table and chair gen under netagaru wants to lie “(When) in this house, our dog is always wanting to lie under tables and chairs” b. * [ kono ie-de uti no inu-ga itumo netagaru ] this house-loc us gen dog-nom always wants to lie tukue ya isu no sita table and chair gen under-loc (lit.) “under tables and chairs, where our dog is always wanting to lie when in this house”

43

Here again, however, the system is not intended to be able to make grammaticality judgements, and given an input such as (22b), kono-ie is simply assumed to ?ll the Inner Locative role. IF (locative head) IF (Locative case slot uninstantiated) RETURN Inner Locative; ELSE RETURN Outer Locative; IF (uninstantiated Subject case slot) RETURN Subject;

4.5.6

Tool-aided action

Tool-aided verbs are implicitly associated with a well-de?ned set of Instrumental which facilitate the action (a ‘tool-set’). Instrumental denotations are verb-speci?c, and Instrumental mentions occur in the locative case. One test for tool-aided actions is the recoverability of a tool sense for a generic object-referring noun head such as mono “thing” or yatu “thing”, when modi?ed by a relative clause containing simple non-past in?ection on the main verb and no overt Subject. (23) [ hasamu ] yatu grabs thing “a grabbing thing/implement” (lit. “a thing (with which) to grab”)

Examples of tool-aided actions are ut(-u) “to strike/hit” (the associated tool-set includes such entries as batto “bat” and kanaduti “hammer”) and kak(-u) “to write” (tool-set includes pen “pen” and fude “brush”). IF (head ∈ verb-defined tool-set) RETURN Instrumental; ELSE IF (generic object-referring head AND simple non-past main verb tense AND relative clause subjectless) RETURN Instrumental;

(i)

kono-ie-de uti-no-inu-ga itumo sita-de netagaru this house-loc us gen dog-nom always under-loc wants to lie tukue-ya-isu table and chair (lit.) “tables and chairs, under which our dog is always wanting to lie when in this house”

44

CHAPTER 4. VERB CLASS-BASED RESOLUTION

4.6

Quotative

Quotative verbs are compatible with clause quotative (subordinating) usages, with the subordinated clause marked with the quotative (to) case marker. When nominalised, quotative verbs can generally express indirect and direct quotation through message linking with the to-no complex case marker or to-i(-u) relational verb.8 For quotative verbs, case-role gapping can occur from within the subordinate clause (“subordinate clause gapping” - see below). Examples of quotative verbs are i(-u) “to say”, omo(-u) “to think” and tutae(-ru) “to report”.

4.6.1

Subordinate clause gapping

For quotative verbs, case-role gapping can extend across a ‘bridging’ clause to a subordinated clause. Bridging clauses are de?ned as accommodating the subordinate gapping process without containing a gap themselves.9 Bridging clauses must be headed by a quotative main verb, and rely on the subordinate clause being marked either with the ‘quotative’ case marker (to), in which case the clause takes ?nite in?ection, or the nominative case marker for nominalised subordinate clauses (koto or no nominaliser). (24) [[ 100-ton izy? o aru ] to mi-rare-ru 100.tonnes over to be-pres quot consider-pass-pres ‘stock considered to be over 100 tonnes (in quantity)’ ziken-ni kanyosi-ta ] to nihon-ga incident-dat contribute-past quot Japan-nom miteiru ] kunii considers country ‘countries which Japan believes to have contributed to the incident’ kaigi-ni sanka-suru ] koto-ga kakunin-sareteiru meeting-dat attends -nml-nom is con?rmed “people who are con?rmed to attend the meeting” [[ ] hito person ] zaikoi stock

(25)

[[

(26)

In order for subordinate gapping to occur for quoted subordinate clauses, the main verb in the superordinate relative clause must be potential or passive, or alternatively the superordinate relative clause must contain a surface representation of the clause Subject; in the case of a nominalised subordinate clause, the main verb in the superordinate clause must be passive. If these in?ectional/syntactic requirements are met, gapping resolution takes place at the subordinate clause level, based on the valency frame and in?ectional content of the subordinate main verb. Interestingly, the same scope of gap types exists at the subordinate level as at the matrix relative clause level. We can thus apply the proposed matrix clause rules unchanged, excepting that subordinate gapping can only occur across a single bridging clause and hence recursion must be limited to a depth of one. If a case-role gap is detected within the subordinate clause, the system returns not only the identity of the gapped case slot, but also the fact that the gap is subordinate rather than superordinate. For quoted subordinate clauses, a passive main verb produces coindexing between the superordinate and subordinate clause Subject positions, and an active main verb leads to the subordinate Subject being coreferent with the superordinate Direct Object.
8 As mentioned in Section 2.1, the frequently occurring relational verb to-i(-u) is excluded from consideration in this research, and hence despite its compatibility with the quotative case marker, it is not classi?ed as a quotative verb. 9 Naturally, if the subordinate clause is to contain a case-role gap, an alternate gap cannot exist within the superordinate clause, according to the ‘gap uniqueness’ maxim.

4.6. QUOTATIVE (27) (sono) kuni (no koto)-o Nihon-ga [ ziken-ni kanyosi-ta (that) country (gen -nml)-acc Japan-nom incident-dat contribute-past to mite-iru quot considers “Japan considers (that) country to have contributed to the incident.” ]

45

These sub/superordinate case slot correspondences are evoked in marking case slot incompatibility at a given clause level, in the case of case slot instantiation at the other level. For example, if the superordinate Direct Object case slot were instantiated in (27), the subordinate Subject case slot would be excluded from the case-slot gapping candidate set for gapping resolution at the subordinate clause level, and vice versa. While recognising that these inter-clausal case slot correspondences super?cially contradict our stipulation that gapping occurs from a unique case slot in a given interpretation, we consider the coindexed case slots to have been con?ated into one, and analyse the gap as existing in the subordinate clause. Indeed, the only consideration of the corresponding superordinate case slots comes in checking for zero content during gapping resolution, and conversely, for stipulating local gapping incompatibility in the case of instantiation of either of the case slots in question. Full con?ation of the superordinate and subordinate case slots can be seen through the unacceptability of a co-instantiation of the corresponding case slots in a matrix clause setting. (28) * (sono) kuni (no koto)-o Nihon-ga [ (sono) kuni-ga ziken-ni (that) country (gen -nml)-acc Japan-nom (that) country-nom incident-dat kanyosi-ta ] to mite-iru contribute-past quot considers “Japan considers (that) country to have contributed to the incident.” (intended)

That the gap exists at the subordinate clause level, rather than the superordinate clause level, can be veri?ed by application of the subject honori?cation test on the subordinate and superordinate clauses. (29) a. kaigi-ni deta ] to sareteiru ] sensei meeting-dat attended quot is said teacher “teachers who are said to have attended the meeting” b. [[ kaigi-ni o-de-ni natta ] to sareteiru ] sensei (Subordinate subject honori?c form of a.) c. * [[ kaigi-ni deta ] to o-sare-ni natteiru ] sensei (Superordinate subject honori?c form of a.) [[

The grammaticality of (29b) over (29c) clearly supports the proposed subordinate case-role gapping treatment. IF (passive or potential main verb OR superordinate Subject position instantiated) Mark any subordinate gap incompatibilities based on superordinate case content; IF (gapping resolution of the subordinate clause identifies a gap α) RETURN SUB-α; ELSE RETURN Content; ELSE mark any superordinate gap incompatibilities based on subordinate case content;

Chapter 5

Miscellaneous processing
5.1 Non-gapping expressions

Non-gapping expressions are de?ned as noun heads which are generally associated with a head restrictive interpretation of the containing relative clause complex. One example of a non-gapping expression is mokuteki “purpose” in (1) below. (1) hataraku ] mokuteki works purpose “the purpose for working” [

Within the current system, detection of a full non-gapping expression head1 is taken to automatically produce a head restrictive clause sense. The fallacy of this strategy and potential for non-gapping expressions to be involved in gapping interpretations, is evidenced in (2). (2) [ kare-ga nobeta ] he-nom gave “the purpose he gave” mokuteki purpose

Despite this realisation of the inherent limitations of the proposed analysis method, non-gapping expressions provide a low-cost and remarkably e?ective means of ?ltering o? head restrictive relative clauses, with the overall bene?t derived through their use far outweighing the inherent noise they produce in analysis. Eventually, the system accuracy is hoped to be brought up to a level equivalent to the performance gain availed by non-gapping expressions, at which point they will lose their worth and a more thorough, context-dependent means of di?erentiating between gapping and head restrictive clauses will become necessary. Examples of non-gapping expressions are: mokuteki “purpose”, ugoki “movement/trend”, h? osin “direction/trend”, kanzi “feeling”, zizitu “fact/truth”

IF (full non-gapping expression head) RETURN Content;

5.1.1

The extraction of non-gapping expressions

Ideally, non-gapping expressions could be mechanically extracted from an annotated corpus, by imposing a threshold likelihood of participation in head restrictive senses, and testing all heads against
1

Note that non-gapping expressions must constitute the full noun head to be able to guarantee non-gapping.

46

5.2. TIME-RELATED ADJUNCTS

47

this threshold value. However, given the limited size of the (annotated) corpus currently used, and its closed set nature, this method could not be applied. Instead, nouns were experientially evaluated for propensity to case-role gapping, and those for which case-role gapping was possible only in relatively restricted domains, were included in the non-gapping expression dictionary. A further automatic learning procedure could be applied to learn verb collocations which produce case-role gapping sense for non-gapping expression heads. Here again, however, a richer source of annotated relative clause instances would be required than is currently available.

5.2

Time-related adjuncts

The following is a description of time-related adjunct types. Time-related adjuncts are by default compatible with all verbs and time-related interpretations override any other interpretations, in the case of ambiguity.

5.2.1

Temporal masking

Temporal masking describes the process of representing the temporal extension of a linguistic unit. A ‘temporal vector’ of discrete time units is employed to this end, containing a slot for each of the year, month, day, and time units: Year Month Day Time

Instances of each of these temporal units in the target linguistic unit are marked by ‘switching on’ the corresponding slot in the vector. For example, 1998-nen no 2-gatu “1998 gen February” would result in the vector: 1 1 0 0

Our interest in temporal vectors lies in their application to the analysis of Temporal gapping relative clauses, where an e?ect of concurrent case-role instantiation quite distinct to that for local adjuncts, is produced. Recall that for locatives, we introduced the notion of inner and outer case-roles, which were related through granularity/speci?city, and interrelated at a high semantic level. For temporal caseroles, the reverse is true, in that while there appears to be an inherent limitation of two on the number of temporals which can include in a given context (including the noun head in the case of relative clauses), the only constraint on multiple Temporals is that their temporal vector representations are not permitted to overlap. In computer hardware terms, when the two temporal vectors are logically AND’ed together, the resultant temporal vector must consist of all zeroes. This leads to the following grammaticality judgements for relative clauses: (3) [ 1-gatu-ni kaigi-ga okonawareta ] January-dat meeting-nom was held “days in January on which meetings were held” hi day

22-niti-ni kaigi-ga okonawareta ] hi 22nd-dat meeting-nom was held day (lit.) “days on which meetings were held on the 22nd” We return to consider temporal vectors in the next section.

*[

48

CHAPTER 5. MISCELLANEOUS PROCESSING

5.2.2

Time relative constructions

Time relative constructions are produced either by the noun head being a time relative expression, or through the combination of the head being a time relative complex and particular main verb in?ectional requirements being met. Examples of time relative expressions are: sono-hi “that day”, yokuzitu “the following day”, t? ozitu “that day” These can collocate with any main verb in?ectional type to generate a time relative interpretation. Time relative complexes are generally produced by attaching a post?x to a phrase describing a time span. Instances of time relative complexes are: 1-kagetu-go “one month later”, nan-nitika-mae “a few days before”, 2-okunen-mae “200 million years before” The two a?xes which can collocate with time relative complexes are -go “after” and -mae “before”. For -go, the stem verb must be in the simple past tense to produce a time relative construction, whereas -mae requires the simple present tense. If the head is a time relative complex but tense and aspectual requirements are not met, a Temporal case-role relative clause is produced. The e?ect of the tense and aspect of the stem verb in variously producing a time relative construction and a non-relative temporal construction, is illustrated by: (4) ky? ory? u-ga sunde-ita yaku-2-okunen-mae dinosaurs-nom were living about 200 million years ago “about 200 million years ago, when dinosaurs lived” ky? ory? u-ga sumu yaku-2-okunen-mae dinosaurs-nom to live-pres about 200 million years ago “about 200 million years before dinosaurs lived”

(5)

Relative clause complex (4) constitutes an absolute temporal construction, and hence Temporal caseslot gapping, whereas the simple present tense in (5) leads to the production of a time relative construction. The justi?cation for the characterisation of time relative constructions as being non-gapping lies in the semantic incompatibility that exists between the time relative interpretation of the construction (extra-clausal), and the interpretation produced for Temporal case-role gapping (intra-clausal). This can be seen in the glosses of sentences (4) and (5) above. Additionally, returning to use of temporal vectors, the observed masking e?ects and mutual exclusivity between temporal expressions co-existing in a single clause, are not observed for time relative constructions (cf. the ungrammaticality of (7)): (6) [ 1963-nen-ni Kened? i dait? ory? o-ga ansatu-sareta ] yokutosi 1963-dat Kennedy president-nom was assassinated the next year (lit.) “the year after President Kennedy was assassinated in 1963” ] tosi year

(7)

1963-nen-ni Kened? i dait? ory? o-ga ansatu-sareta 1963-dat Kennedy president-nom was assassinated (lit.) “the year President Kennedy was assassinated in 1963” *[

The temporal vector for both 1963-nen and yokutosi in (6) becomes: 1 0 0 0

5.2. TIME-RELATED ADJUNCTS

49

Hence, in applying one as a mask over the other, the inital ?ag remains set. This provides a type test for time relative expressions, and allows us to draw a distinct line between time relative clauses and Temporal gapping clauses. This is worthy of particular note, as time relative clauses have been largely misrepresented as case-role gapping relative clauses, with gapping occurring from the temporal case-role (Matsumoto 1997:53).

5.2.3

Temporal expressions

Temporal expressions consist of time-related NP heads which are ground either absolutely or situationally. They consist of absolute temporal expressions, generic temporal expressions and non-relative temporal constructions. Absolute temporal expressions are of the type: 16-niti “the 16th”, sakunen “last year”, mainiti “everyday” That is, they constitute the set of temporal expressions which are well de?ned within the context of the surrounding text. Generic temporal expressions are of the type: kikan “period”, zikan “time”, nendo “year/?scal year”, hi “day” These express generic temporal categories and are semantically restricted by the clause body. They can be likened to lambda expressions in that they are ground time-type case slot ‘casts’, without having the semantic extra-clausal and intra-clausal semantic incompatibility described below for time relative constructions. Non-relative temporal constructions are temporal constructions which involve a time relative complex head, but which do not ful?ll the stem verb in?ectional requirements of a time relative construction (see above). Non-relative temporal constructions produce Temporal case-slot gapping relative clause sense. Note that there is a certain degree of reliance on the surrounding context as to whether a temporal expression is absolute or generic, in that most absolute expressions can be forced to take a generic reading. This di?erence is most noteworthy when analysing restrictive and non-restrictive relative clauses, a matter which is beyond the scope of the current research.

5.2.4

Temporal vs. Durational interpretations

We have de?ned two distinct time-type case-roles in our case-role set, and obviously require some means to di?erentiate between them. For certain relative clause complexes, construal is uniquely de?ned, as occurs with non-relative temporal constructions being mapped onto the Temporal caserole. For absolute temporal expressions, the same applies, in that only the only time-type compatibility exists with the Temporal case-slot. However, for generic temporal expressions, the procedure becomes dependent on an array of factors, including the verb sense, verb in?ection, instantiation of a temporal case slot within the relative clause body, and noun head type. These are modelled heuristically as: IF (non-relative temporal construction OR head suffixed durationally) RETURN Durational; ELSE IF (head is ‘‘zikan’’) IF (‘start’ or ‘end’ main verb) RETURN Temporal; ELSE IF (potential inflection on main verb) RETURN Durational;

50

CHAPTER 5. MISCELLANEOUS PROCESSING ELSE IF (temporal case slot instantiated OR ‘locational action’/‘physical movement’/‘existential’ main verb) RETURN Durational; ELSE RETURN Temporal; ELSE IF (generic temporal head) IF (temporal case slot instantiated) RETURN Durational; ELSE RETURN Temporal;

These heuristics are evaluated in the overall evaluation of Section 8.2.

5.3

Cardinal adjuncts

Cardinal adjuncts are treated as being compatible with only the locational action, physical movement and mental action verb classes. IF (cardinal noun head AND locational action/physical movement/mental action main verb) RETURN Cardinal;

5.4

The default rule set

In the event that none of the rules for the di?erent verb classes is triggered to produce a non-Locative analysis type, a default rule set is applied. The ?rst stage of the rule set is to match the case slot contents of the relative clause against the valency frame, and ascertain remaining complement compatibilities. If the Subject case slot is found to be uninstantiated and the noun head to be animate, a Subject interpretation is returned. This treatment is founded on the “humanness hierarchy” proposed by Kuno and Kaburaki (1977)2 , and the realisation that humans and other animate arguments are more highly empathised than inanimate objects, and hence occupy the Subject case slot more easily. Failing this, ?rst person pronoun heads are analysed as producing a Bound Subject relative clause sense, based on the “Speech-act Participant Empathy Hierarchy” of Kuno and Kaburaki (1977). If the Subject case slot is instantiated and the noun head is not in the ?rst person, then we map the noun head onto the most accessible uninstantiated obligatory complement case slot. Accessibility is de?ned through argument status and the linear ordering of the case slots in the case frame.3 For this obligatory complement case slot mapping process to fail, all obligatory case slots in the valency frame must be instantiated, and the noun head cannot be compatible with any middle or adjunct case slot, as they would have been triggered prior to the defualt rule set. The only remaining interpretations for the relative clause, hence, are bound gapping and head restrictive. Abstract heads are assumed to be unavailable to bound gapping, and hence automatically tagged as head restrictive. For non-abstract heads, the system attempts to identify a complement case slot which is instantiated with a common noun (working down the accessibility hierarchy), and if successful, that case slot is returned as containing the bound case-role. IF (uninstantiated Subject case slot)
2 The same preference for animate Subjects has been implemented into Nishida et al.’s (1980) Case-based machine translation system. 3 Case slots in the valency frame are ordered in descending order, from back to front, with optional complements excluded from analysis at this stage.

5.4. THE DEFAULT RULE SET IF (animate head) RETURN Subject; ELSE IF (first person head) RETURN Bound Subject; IF (uninstantiated obligatory complement case slot) THEN (identify most accessible obligatory case slot α) RETURN α; ELSE IF (abstract head) RETURN Content; ELSE (determine highest accessible common noun-filled complement case slot β ) RETURN Bound β

51

Chapter 6

Lexical ambiguity
This chapter describes methods of resolving lexical ambiguity in the main verb and noun head, through statistical/representational preference and thesaurus use, respectively.

6.1

Verb lexical ambiguity

Plurality of successfully parsed entries results from a combination of both full and partial verb homophony and homography. Full verb homophony is a direct result of the existence of multiple inter-replaceable writing systems within Japanese (hiragana, katakana and kanji), and occurs when two distinct verb entries coincide in both conjugational type and phonetic content of the verb stem/auxiliary verb complex. It is distinguishable from polysemy by virtue of the fact that disambiguation is achievable through use of the kanji form of the verb stem. An example of full verb homophony is “a(-u)”, for which three heterogeneous kanji forms produce the distinct entries corresponding to the generic glosses of “to meet” (q &), “to coincide” (g&) and “to encounter” ()&). Full homophony can alternatively be produced through combinations of auxiliary verb morphemes, such that “miau” is ambiguous between mi-a(-u) “to see--mutual” and mia(-u) “to correspond”. Full verb homography is analogous to full verb homophony, except that the ambiguity exists in the kanji-based representation for coinciding conjugational types. In this case, disambiguation is possible through the kana phonetic version of the verb in question. An example of a full homograph occurs for the verbs tome(-ru) “to stopTRANS ” and yame(-ru)to quit/put an end to, for which a common kanji (“ _”) corresponds to the “to-” and “ya-” pre?xes, respectively. Partial verb homophony, meanwhile, occurs for verbs which di?er in conjugational type, but agree in phonetic content of the verb stem. In this case, heteronomy of kana representation is produced for only certain in?ectional types. In the case of our example of a(-u), ar(-u) “to have” shares the verb stem of a-, and a heteronym is produced in the simple past tense, in the form of atta. Here again, however, kanji representation allows us to resolve the lexical ambiguity. Partial verb homography closely resembles partial verb homophony, except that the lexical ambiguity is produced in the kanji form, and resolvable through the use of kana. One example of partial homophony is produced for the simple past tense verbs i-tta “to go--past” and okona-tta “to carry out/hold--past”, in that a single kanji (“T”) is used to represent both “i-” and “ okona-”, respectively. Note that in both of the classi?cations of partial heteronymic correspondence, the degree of coincidence is usually highly restricted, unlike full verb heteronymy. For the i-tta/okona-tta ambiguity, for example, partial heteronymy occurs only in the simple past tense or for progressive/perfective aspect. 52

6.2. RESOLVING VERB LEXICAL AMBIGUITY

53

6.2

Resolving verb lexical ambiguity

The most immediate method of resolving representational ambiguity is through statistical means. In this, two methods of statistical weighting were tested, the ?rst based on naive probability and the second on representational preferences. For both methods, statistical scores were computed only in cases where multiple “generalised” (i.e. non-?xed expression) verb senses existed for a common verb stem. Fixed expression verb senses were automatically allocated a score of one, on the assumption that the ?xed case element content is mutually exclusive for a given verb stem, and that the system should prefer idiomatic verb senses over generalised verb senses in the case that an idiomatic verb sense is plausible. Due to the di?culty in predicting partial verb homophony/homography, all verbs sharing a common stem are treated as being fully heteronymic. Note however that for most in?ectional types, coincidence of in?ectional form does not result. In this case, the preferred verb sense is the one which has the highest relative score, ignoring the fact that the various scores in question may not total to one. In terms of the interface between statistical weighting and the rule set, the rule set is applied as is for each parseable verb entry, and weights are summed for each resultant output. The unique system output is determined simply by calculating the highest summed weight, and randomly selecting between multiple analysis types of highest score.

6.2.1

Calculation of verb scores

The collation of frequencies is based on the EDR corpus (EDR 1995), and the verb sense annotations given for each verb occurrence. This is the same corpus as was used to extract all relative clause test sets described in this paper, and hence forms a closed test set. Whereas no direct reliance is made on verb sense by our system, the EDR corpus provides a means of determining lexical correspondences between di?erent verb forms. To return to our example of atta above, all occurrences of atta are attributed a verb sense index, which correspond to di?erent verb ‘sense sets’. Contained in these sense sets are one or more representational alternatives of the verb stem a-, detectable through the verb stem representation contained in the original system dictionary entry. While there is not guarantee of disjunction between the alternative forms of atta and their respective sense sets, in almost all cases seen in the EDR corpus, full disambiguation was possible through the granularity of the verb sense index. In cases where sense ambiguity remained, the frequency of the original verb index was equally distributed between polysemous candidates. One unfortunate characteristic of the EDR corpus is the uncommonly high numbers of index mismatches and ‘nil’ verb senses (unanalysed/unanalyseable verb senses). In the calculation of verb scores, index mismatches were simply disregarded from the data, while ‘nil’ indices were treated as described below for the separate scoring methods. Frequencies are calculated a priori and normalised (signi?cant to three ?gures) to produce the probability of occurrence of that form of the given stem verb.

Naive probability of occurrence
The naive probability of occurrence (NPO ) of lexical form a of verb entry f (represented as af ) is computed simply by totalling the number of usages of verb senses corresponding to af , and normalising over the total occurrences of a. Smoothing is achieved by evenly distributing ‘nil’ occurrences for a between entries ai , where the total number of distinct entries ai is represented as |a| in equation (6.1). Thus, high levels of ‘nil’ occurrences will produce roughly standardised probabilities for all entries ai , whereas lower levels of ‘nil’ occurrences will lead to nearer correspondence between relative frequency and normalised probability. This is intended to re?ect the assumption that ‘nil’ senses suggest inherent ambiguity, and that higher levels of ‘nil’ values indicate lower con?dence on the part

54 of the EDR developers in annotating usages of a. N P O(af ) =
f req (ani l ) +f req (af ) |a| i

CHAPTER 6. LEXICAL AMBIGUITY

f req (ai )

(6.1)

Normalised representational preference
The representational preference (RP ) of lexical form a of verb entry f (i.e. af ) is de?ned as the con?dence with which one can predict that a will be used to represent f , with the mean con?dence predicted as 1. Smoothing is carried out through a double application of Je?rey’s estimate (Good 1965), that is by adding one to both the numerator and denominator. In this way, low-frequency verb entries and lexical forms can be smoothed to a value near the mean con?dence of one (or to exactly one for zero-frequency entries), but at the same time high-frequency items are relatively una?ected. Additionally, instances of zero denominators are avoided, and the con?dence is guaranteed to be strictly greater than zero. Occurrences of the ‘nil’ index are not included in the RP calculation, such that entries found only with the ‘nil’ index return a representational preference of one. RP (af ) =
1+ 1+f req (af ) f req (if ) i=a

(6.2)

This is normalised over the representational preference for all source entries ai , to produce the normalised representational preference N RP (af ). N RP (af ) =
RP (af ) RP (ai ) i

(6.3)

6.2.2

Complexity of in?ectional content

The only representational ambiguity not covered by these two scoring systems is instances where in?ectional morphemes have produced an ambiguity which was not predictable from the stem verb (see the example of miau in section 6.1). This shortcoming is resolved by introducing the concept of ‘complexity of in?ectional content’ (CIC ), in which we penalise higher numbers of component in?ectional morphemes. The penalty is computed in situ based on the number of in?ectional morphemes contained in the verb, relative to the parse of simplest in?ectional content (min in? ); the simplest parse receives a complexity of one. Thus, in the case of “miau”, mia-u “to correspond-pres” has a complexity of one, and mi-a-u “to see-mutual-pres” has a complexity of two. Weighting is achieved through the use of the constant parameter α. That is, the relative contribution of CIC can be enhanced by increasing α, hence exponentially increasing the value of the denominator and reducing the overall verb score (VS ). At the same time, the parse of simplest in?ectional content receives a complexity of one, and its VS is hence una?ected by variation in the value of α. Complexity of in?ectional content is compatible with both methods of statistical weighting given above, such that the VS for lexical form a of entry f (i.e. af ) using statistical weighting measure SW is computed by: V S (af ) =
SW (af ) (CIC (af )?min inf l+1)α

(6.4)

6.2.3

Evaluation of verb scoring

Preliminary evaluation was carried out to determine the relative e?ectiveness of the naive probability of occurrence (NPO) and normalised representational preference (NRP) methods, and contribution of CIC. The test sets used for this purpose were the full set of annotated relative clauses used in

6.3. NOUN HEAD LEXICAL AMBIGUITY Overall accuracy (4411) Baseline NPO (α = 1) NPO (α = 10) NRP (α = 0) NRP (α = 1) NRP (α = 10) Optimal 84.6% 86.0% 86.0% 85.9% 85.8% 85.9% 88.4% Accuracy on case-role gapping clause instances (3650) 90.8% 92.3% 92.3% 92.1% 92.2% 92.2% 94.5%

55

Table 6.1: Results for the verb scoring methods

developing the system, and the subset of gapping relative clauses. The sizes of the two test sets are indicated in brackets below each heading. The baseline method for evaluation purposes simply selects the verb sense of highest probability when multiple parses are produced, which equates to utilising the naive probability method in computing the verb score, with α set to zero. The optimal achievable result for the system is determined by testing for membership of the correct analysis in the full set of analysis types produced for all successful parses. Given that verb scores simply rank these candidates, it is impossible for the other methods to better this non-deterministic method. Table 6.1 lists the comparative results for the various methods1 , including evaluation of varying values of α for both the NPO and NRP methods. The 1.4% point di?erence between the overall accuracy for the baseline method and that for the NPO method with various values of α is a direct indication of the e?ects of weighting according to in?ectional complexity, although the ine?ectiveness of an increased α value is unexpected. Likewise for the NRP method, whereas results are signi?cantly higher than those for the baseline method, altering α produced only minor improvement. Indeed, performance with α set to zero (i.e. without consideration of CIC) marginally outperformed NRP with α set to one, although the statistical signi?cance of this di?erence is questionable. This would tend to suggest that there is some interference in the choice of representational form of the verb stem given complex in?ection, a fact which was borne out on summary inspection of the data. That is, the kanji form of the verb stem is generally utilised if auxiliary verbs are also given in a kanji representation, and full hiragana representation is generally reserved for simple in?ection uses, such that a hiragana occurrence of “miau” would tend to point to the simple in?ectional ‘mia-u’ stem (see section 6.1). Perhaps more noticeable, however, is that the NPO method slightly outperforms NRP, which leads to the conclusion that representational preference in isolation is outweighed by the brute force of likelihood of sense. Based on these results, we adopt the NPO method for the remainder of this paper, with α set to one.

6.3

Noun head lexical ambiguity

Noun head lexical ambiguity arises because of the polysemous representation utilised in the NTT thesaurus (Ikehara et al. 1993). That is, if a word displays polysemy, its various senses are positioned
1

Data taken from (Baldwin 1998).

56

CHAPTER 6. LEXICAL AMBIGUITY

separately in the thesaurus, rather than attempting to maintain a one-to-one isomorphism between lexical form and thesaurus correspondences. This presents us with a dilemna, as we wish to not only classify lexical arguments according to type (i.e. as ‘animate’, ‘locative’, etc.), but also to weight the di?erent senses so as to be able to chose between adjunct and complement senses, for example. The method we use to weight nouns (W (N )) on class typicality, is simply to count the total number of occurrences of that noun N in the thesaurus, and the number of occurrences which fall into the particular sub-trees we have designated as classifying a particular type T , and calculate the ratio thereof. W (N ) =
f req (N ∈T ) f req (N )

(6.5)

For a noun N which never occurs in the extension of T , W (N ) thus becomes zero, whereas for an N fully enclosed within T , W (N ) is one. In terms of the application of this weighting scheme to class membership, we stipulate a threshold for animacy, such that W (N ) must be greater than or equal to 0.5 for N to be judged as animate. For the locative class, on the other hand, we skew the distribution of the produced W (N ) to produce preference for highly prototypical locatives, over animacy and other judgements, but penalise less clear-cut examples. The way we do this is to apply the function Loc(N ): Loc(N ) =
[W (N )+1]5 21

(6.6)

What this crude and computationally expensive function does is to in?ate values closer √ to one, to a maximum of around 1.52, and penalise anything under a value of around 0.84 (actually, 5 21 ? 1) by way of a relatively steep parabolic curve (values near zero actually increase slightly).

Chapter 7

Extensions to the basic algorithm
Despite the obvious attractions of the algorithm in the form presented to here, and its ability to weight interpretations, it still lacks in its ability to capture inter-clausal context, in what turns out to provide a surprisingly rich source of restrictions on the interpretation type. Here, we discuss the processing of cosubordinated clauses, coordinated clauses and coordinated heads.

7.1

Relative clause cosubordination

Clause cosubordination in Japanese is indicated by the use of a cosubordinating conjunction of the type nagara, te, tutu and si, or through ren’yo type in?ection (aka. continuative (Kuno 1973b)). (1) [[ pasukaru-ga ti k? oan-si, ] ti seisaku-si-ta Pascal-nom DO design-ren DO make-past ‘a computing device designed and produced by Pascal’ [[ ti arubaito-o si-nagara ] SBJ part-time work-acc to do-while ‘students who work part-time while at school’ ti SBJ ] keisan-kikaii computing device

(2)

gakk? o-ni school-dat

kayo-u attend-pres

]

gakuseii student

Within cosubordinating connectives, Kuno (1973b) observes that si and ren’yo must be subject coreferential, and Yoshimoto (1986) and Minami (1974) note that all cosubordinating connectives tend to coincide in Subject or Object content. These observations are borne out for (1) and (2). In terms of relative clause analysis, we wish to suggest (3) as a corollary of the mutual exclusivity of the gapping paradigm: (3) All component cosubordinated and subordinated clauses within a complex relative clause must agree in relative clause type.

That is, it is not possible to have a relative clause comprised of both gapping and non-gapping clause components. Additionally, we extend the observations of the above researchers to hypothesise that: (4) For cosubordinated gapping relative clauses, the component clauses must agree in case-role gap type.1

In this, we wish to distance ourselves from peripheral adverbial usages of nagara and tutu (in which the nagara/tutu su?x is interchangeable with nagaramo in the contrastive sense and toki in the manner sense) and non-additive usages of te (Hasegawa 1996:6). Note that as was the case for subordinate gapping, the scope of gapping is unrestricted between complement and adjunct case slots, and includes, in this case, subordinate gaps.
1

Note that this coincidence of gap does not apply to anchored clauses.

57

58

CHAPTER 7. EXTENSIONS TO THE BASIC ALGORITHM

7.1.1

Processing of clause cosubordination

By way of accepting hypothesis (4) on gap type correspondence, we are able to extend our algorithm to consider case slot incompatibilities, in addition to the existing framework of case slot compatibility determination. Case slot incompatibilities stem from two sources: (i) directly from the content of the complement valency frame, and (ii) from case slot instantiation. Given a tool set of complement caserole types, it is possible to determine inherent case incompatibilities directly from the valency frame of the verb in question through a simple matching mechanism. This is combined with an analysis of those case slots instantiated in the input, and hence incompatible with that gap through the ‘one-caseper-clause’ constraint (Fillmore 1968:22).2 Given that we can expect multiplicity of analysis type due to multiple parses, we take the intersection of gap incompatibilities for each analysis type, and return the resultant set of incompatibilities for the highest scoring analysis type. On the inter-clausal level, the union is taken of the individual incompatibility set for each component clause, in determining the overall incompatibility set. Determination of the unique overall analysis for the relative clause is facilitated through the same process as at the single clause level, in that the weighted outputs for each member clause are summed, and a ?nal sorted list of analysis types determined. However, this is now combined with the incompatibility set to weed out incompatible case-role types, and the highest scoring compatible clause analysis is outputted. In the case that all analysis types are judged to be case incompatible, the overall clause is assumed to be non-gapping.

7.1.2

Gap correspondences

Cosubordination of canonical gapping and subordinated gapping clauses leads to an interesting e?ect, in that inter-clausal agreement occurs in terms of the gap type, but not as to the clause level from which gapping has occurred (see (5)). It is for this reason that our hypothesis stipulates agreement in case-role type sub-type, but makes no mention of clause level. (5) ti i-na-i ] to mi-rare ] ti SBJ to be-neg-pres quot consider-pass-ren SBJ renraku-sare-na-katta ] hitoi to contact-pass-neg-past person ‘a person who was assumed not to be in and (hence) not contacted’ kankeisya-o nozo-ki ] ti p? ati-ni syussekisi-ta organiser-acc exclude-ren SBJ party-dat attend-past ‘the number of people who attended the party, excluding organisers’ [ yoru [ t? oky? owan-o watari-nagara ] ti mi-ru ] night Tokyo Bay-acc to cross-while DO reinb? oburizzii Rainbow Bridge ‘Rainbow Bridge as seen at night while crossing Tokyo Bay’ [[ [[ ] ninz? ui number of people [[[

(6)

(7)

to see-pres

(8)

Tar? o-ga sekininsya-to nari, ] 1-gatu-ni Taro-nom person in charge-com to become-pres January-dat okona-ware-ru ] konk? uru to hold-pass-pres competition ‘a competition to be held in January, which Taro is in charge of organising’

2

Note that application of the one-case-per-clause constraint is restricted to complement case slots.

7.2. COORDINATED RELATIVE CLAUSES

59

Two verb types which do not contribute to the clause sub-type, and are hence disregarded during the resolution process, are the excluding and including types. Excluding and including clauses are adverbial constructions, and hence exempted from consideration with hypothesis (4). Considering (6), in which the ?rst clause is of the excluding type, the main clause is essentially treated as a simplex clause, and the Subject gapping sub-type can be recovered. One fact which is clear from the original description of conjunction types is that peripheral subordinating usages exist for all conjunctions except the ren’yo form, suggesting di?culty in correctly predicting the type of clause dependency in a given clause prior to being able to apply the restrictions proposed in section 7.1.1. While this is certainly the case for te clauses, complement analysis-based heuristics were found to be productive in correctly analysing nagara and tutu clauses. These heuristics consist of analysing the complement content of the cosubordinated clause to determine if all nonSubject complement case slots are instantiated. The exceptional treatment of the Subject case is founded in the observation that these are ‘small clauses’ (Radford 1981), the Subject of these subordinate clause types is inherently coindexed to that of the superordinate clause, through a PRO mechanism, and overt Subject mention within the nagara clause is not possible. If full instantiation is detected, the unit clause in question is therefore discounted from the resolution process, on the grounds of being adverbial. This process can be seen to correctly identify the subordinated nagara clause in (7), with the Direct Object gap existing only in the main clause and no Direct Object incompatibility restriction imposed by the nagara clause. One additional quali?cation which must be made to (4) is that it does not seem to apply to the bounded case-role for bounded relative clauses. To take an example, the Direct Object case-slot is bound in the ?rst clause of (8), but the ?nal clause is a clear instance of Subject case-slot gapping. At the present time, we have no explanation for this e?ect, and simply disregard bounded relative clauses during cosubordination-based resolution. Note, however, that this does not threaten the applicability of (4), as bounded relative clauses are included under the classi?cation of case-role gapping relative clauses.

7.1.3

The treatment of subordinate clauses

A preliminary study of the relative clause corpus produced for the system suggested that around 6–7% of all relative clauses involve clause cosubordination, pointing to the signi?cance of the above method of analysing cosubordinate clause complexes. Relative clauses containing adverbial clauses, however, seem to account for a much higher proportion, at around 20% of all relative clauses. While the clause sub-type hypothesis proposed above does not apply to adverbial relative clauses, the more general suggestion of coincidence of relative clause type (gapping vs. head restrictive) is suggested to apply to all relative clauses. As is evident in all levels of evaluation, the accuracy of the system for head restrictive clauses is signi?cantly lower than that for gapping clauses, and the application of this basic restriction presents itself as a possible tool in enhancing resolution of the clause type. One promising starting point for an extension of this type comes from the work of (Okumura and Tamura 1996:874), who suggest that ‘subject switching’ occurs between adverbial and containing clauses, given a surface Subject in either clause. They then go on to comment that variations in gap type are largely context dependent and not predictable simply from local constraints. The application of their proposed heuristic, and further analysis of the gap switching mechanism, however, remain as outstanding issues in the system development.

7.2

Coordinated relative clauses

Coordinated relative clauses are essentially multiple instances of ?nite-in?ected relative clause bodies, which share a common head. They do not adhere to the stronger-version hypothesis (4), as seen below, but we hypothesise that the weaker hypothesis of (3) is maintained.

60 (9) [

CHAPTER 7. EXTENSIONS TO THE BASIC ALGORITHM ano y? umeikyoku-o sakkyoku-sita, ] [ daremo-ga that famous piece-acc composed everyone-nom yoku sitte-iru ] B? et? oben well knows Beethoven “Beethoven, who you all know well, and wrote that famous piece of music”

The potential for coordinated case-role gapping relative clauses to be associated with discrete caserole gaps, is illustrated with (9), for which the ?rst clause in Subject case-role gapping, and the second Direct Object case role gapping. Due to this possibility for non-case-role correspondence, it is necessary to annotate and process coordinated relative clauses independently. Currently, however, relative clauses are attributed a unique overall interpretation, making annotation of case-role switching coordinated relative clauses impossible.

7.3

Noun head coordination

A more approachable problem is the handling of noun head coordination. Unlike relative clause coordination, noun head coordination occurs within the noun head, alleviating scope for parallel interpretations for separate noun units. This means that we can safely process each noun head in parallel and combine the resultant analyses to determine the analysis type of overall highest plausibility. In this, combination of interpretational scores is performed additively.

Chapter 8

Evaluation
8.1 Evaluation criteria

Evaluation of the system was carried out based on a test set of relative clause complexes extracted from the Japanese EDR corpus (EDR 1995). The test set was classi?ed according to the verb class content of the main verb, to verify the accuracy of each verb class type. Additionally, ?xed expressions were identi?ed to compare the overall system performance based on generalised valency frames (Table 8.1) , and that for ?xed expressions (Table 8.2). The system accuracy was analysed according to the overall accuracy for each test set, and also the accuracy on only case-role gapping examples1 . For each class of rule set outputs given below, the actual number of instances of that class is given in “Total”, followed by the overall number of correctly and incorrectly analysed relative clauses. The data is then broken down into the case-role gapping and head restrictive types, following which an analysis is given for each clause sub-classi?cation receiving an output for that verb class. The ?gures are additionally each analysed in terms of precision, recall and F-measure (β set to 1 throughout evaluation). # of instances of that type correctly identi?ed by the algorithm Total # of instances categorised as that type by the algorithm # of instances of that type correctly identi?ed by the algorithm Total # of instances of that type (β 2 + 1) × Recall × Precision β 2 × Recall + Precision

Precision = Recall = F-measure =

Cases where a zero denominator has made any of these values incalculable are indicated in the results as ‘N.C.’ (Non-Calculable). Evaluation of total performance for the given data set is calculated only in terms of precision (accuracy), but this ?gure is identical to that for the system recall on the full input set.

8.1.1

Baseline evaluation

As in any empirical evaluation, it is vital that we establish a baseline ?gure for the system performance. As summary analysis of the case-role gapping correspondences for the overall data set show, the Subjectcase-role alone accounts for 3004/4615 (= 65.09%) of all relative clause instances. Thus, by establishing an algorithm which automotically outputs a Subjectanalysis for any input, irrespective
In terms of the tags returned from the algorithm, all instances of Exclusive, Inclusive, Contentand Time Relationalwere excluded from the data.
1

61

62

CHAPTER 8. EVALUATION

of head type or verb class membership, we are able to attain an overall accuracy of 65.09%. This forms the true baseline performance for our system (B1 ). An alternative baseline performance ?gure can be obatined from the implementation of the algorithm proposed in (Baldwin et al. 1997a), which constitutes a much simpli?ed version of our ?nal system, but relies on the same basic concepts and methods for non-gapping expressions, Temporal case gapping, time relative constructions, and and case slot instantiation. First, non-gapping expression-headed relative clauses are ?ltered o? as generating head restrictive relative clauses. Next, the system accesses a transitivity judgements for the main verb of the input realtive clause, based on which the system attempts to map the head onto the Direct Object case slot (assumed accusative case marking) for transitive verbs, and the Subject case slot (assumed nominative case marking) failing this. As a default, all relative clauses are assumed to be head restrictive. This algorithm (B2 ) produces signi?cant improvement over B1 above, with an overall accuracy of 75.1%, and is detailed along with B1 in Table 8.1.

8.2

Overall evaluation
# instances Total Baseline (B1 ) Baseline (B2 ) Case-role gapping Head restrictive Bound Subject Subject Direct Object Indirect Object Co-actor Co-patient Medium Target Passive Agent Causee Instrumental Locative Cardinal Allative Perlative Temporal Durational Exclusive Inclusive Time Relational Idioms Subordinate gapping 4615 4615 4615 3913 601 51 3004 306 15 62 1 18 7 1 0 1 99 16 23 4 97 49 114 28 19 82 17 # correct 4089 3004 3466 3659 330 33 2890 270 13 52 1 9 5 0 0 1 70 8 17 3 93 46 113 28 19 81 17 # incorrect 526 1725 1149 481 45 32 262 93 4 8 0 2 1 0 1 1 41 8 5 1 7 8 2 4 0 0 1 Precision 88.60 65.09 75.10 88.38 88.00 50.77 91.69 74.38 76.47 86.67 100.00 81.82 83.33 N.C. 0.00 50.00 63.06 50.00 77.27 75.00 93.00 85.19 98.26 87.50 100.00 100.00 94.44 Recall – – – 93.51 54.91 64.71 96.21 88.24 86.67 83.87 100.00 50.00 71.43 0.00 N.C. 100.00 70.71 50.00 73.91 75.00 95.88 93.88 99.12 100.00 100.00 98.78 100.00 F-measure – – – 90.87 67.62 56.90 93.89 80.72 81.25 85.25 100.00 62.07 76.92 N.C. N.C. 66.67 66.67 50.00 75.56 75.00 94.42 89.32 98.69 93.33 100.00 99.39 97.14

Table 8.1: Overall analysis

The overall performance of the system on the corpus of 4615 annotated relative clauses is detailed in Table 8.1. The overall system accuracy calculates to around 89%, as compared to 65% for the true baseline Subject case-slot analysis method (B1 ) and 75% for the naive transitivity algorithm (B2 ).

8.2. OVERALL EVALUATION

63

Within the ?gure of 89%, the contribution from the case-role gapping and head restrictive relative clause groups is approximately equivalent. The ?rst major result is the disparity between the recall for these two relative clause types, with case-role gapping clauses far outperforming head restrictive clauses on 95%, as compared to 55%. This points to there being an over-bias towards case-role gapping gapping clause analysis, and overgeneration occurring for this type, an unsurprising result given the core focus of this research on case-role gapping clauses. Within the case-role gapping relative clause type, however, the ?gures for bounded relative clauses are slightly disappointing, and again there are signs of overgeneration occurring. An interesting correspondence between case-slot accessibility/immediacy and accuracy, for the complement case-role set, with performance gradually degrading from Subject to Direct Object, Co-actor and Indirect Object. It must be said, however, that this trend is probably not entirely coincidental, due to the focus placed on the more accessible case-roles during the verb class production phase. The worst ?gures are seen for the Local case-role set, a sign of the frequent ambiguity between the locality and animacy/autonomy senses, as occurs for country name references. Additionally, the context-independence of locative detection leads to the system occassionally missing the local sense altogether. Having said this, it is reassuring to note that the lowest F-measure values are at least comparable to the accuracy for the true baseline of B1 , with the Allative and Perlative gapping analyses roughly equivalent in degree to B2 . The treatment of the time case-slots and time relatives was, if anything, better than expected, and the main source of noise between the Temporal and Durational case-slots was mistaken mapping between the two. That is, the system is generally able to ascertain the time-relatedness of time case-gapping, but has slight di?culties in di?erentiating between the two sub-types. Similarly, the system performed remarkably well for the well-de?ned head restrictive sub-classifications, with only the Inclusive sense falling below a 95% F-measure value. One conclusion which could be drawn from this is that the successful handling of other well-de?ned head restricting phenomena could well be the most e?cient method of further improving system performance, and eliminating overgeneration of case-slot gapping interpretations. Finally, full clause-based idiom detection was predictably excellent, as was the identi?cation of subordinate gapping instances.

8.2.1

Evaluation of ?xed expressions

Fixed expressions were singled out for separate evaluation due to their high presence in the system dictionary. Due to the slightly lower average number of generalised case slots within a ?xed valency frame, as compared to a (fully) generalised valency frame, one would expect them to perform signi?cantly better than their generalised counterparts. The ?gures in Table 8.2 do not support this supposition, however, principally because of the surprisingly low precision and recall for head restrictive relatice clauses. As a means of comparison, we decided to compare the performance of ?xed expressions-based analysis for ?xed expressions, with a generalised valency analysis, to test whether the relatively low accuracy represents some inherent complexity in the data set, or simply an erroneous analysis procedure. The results for this generalised analysis are given in the bracketed “Generalised” row. Happily, there is a signi?cant performance gap between the teo analysis methods, with ?xed expression-based processing gaining a 4 percentage point advantage over generalised analysis. One additional statistic extracted for the ?xed expression data set was the number of occurrences of head displacement, and the accuracy of ?xed argument matching. As indicated in the “Head displacement” row, the system was able to detect ?xed argument displacement with 100% precision and recall. Given that the di?erence between the generalised and ?xed expression analyses was within the order of these 16 occurrences of head gapping, we then veri?ed the accuracy of generalised analysis on these 16 instances. This revealed that the generalised analysis method was able to correctly reproduce

64
# instances Total (Generalised) Head displacement Case-role gapping Head restrictive Bound Subject Subject Direct Object Indirect Object Medium Locative Cardinal Allative Temporal Durational Inclusive Time Relational 212 212 16 182 29 2 136 29 0 3 1 2 4 1 1 1 1 # correct 183 (174) 16 173 9 0 133 28 0 3 1 0 4 1 0 1 1 # incorrect 29 (38) 0 26 3 1 20 3 1 0 0 0 0 1 0 0 0

CHAPTER 8. EVALUATION
Precision 86.32 (82.08) 100.00 86.93 75.00 0.00 86.93 90.32 0.00 100.00 100.00 N.C. 100.00 50.00 N.C. 100.00 100.00 Recall – – 100.00 95.05 31.03 0.00 97.79 96.55 N.C. 100.00 100.00 0.00 100.00 100.00 0.00 100.00 100.00 F-measure – – 100.00 90.81 43.90 N.C. 92.04 93.33 N.C. 100.00 100.00 N.C. 100.00 66.67 N.C. 100.00 100.00

Table 8.2: Analysis of ?xed expressions

# instances Total Case-role gapping Head restricting Subject Direct Object Locative Temporal Subordinate gapping 50 44 6 37 4 1 2 15

# correct 45 (26) 44 (25) 1 37 4 1 2 15

# incorrect 5 (24) 5 (24) 0 2 2 1 0 1

Precision 90.00 (52.00) 89.80 (51.02) 100.00 94.87 66.67 50.00 100.00 93.75

Recall – – 100.00 (56.82) 16.67 100.00 100.00 100.00 100.00 100.00

F-measure – – 94.62 (53.76) 28.57 97.37 80.00 66.67 100.00 96.77

Table 8.3: Results of subordinate gapping analysis

14 of the 16 displacement instances,

赞助商链接

更多相关文章:
更多相关标签:

All rights reserved Powered by 甜梦文库 9512.net

copyright ©right 2010-2021。
甜梦文库内容来自网络,如有侵犯请联系客服。zhit325@126.com|网站地图