Japanese Sentence Processing

Reiko Mazuka, Noriko Nagai, Reiko Mazuka, Noriko Nagai

• Publisher: Psychology Press
• Year: 2013
• ISBN:
  9781134770212
• Edition: 1
• Topic: Persönliche Entwicklung
• Subtopic: Schreib- & Präsentationsfähigkeiten

This volume is a direct result of the International Symposium on Japanese Sentence Processing held at Duke University. The symposium provided the first opportunity for researchers in three disciplinary areas from both Japan and the United States to participate in a conference where they could discuss issues concerning Japanese syntactic processing. The goals of the symposium were three-fold: * to illuminate the mechanisms of Japanese sentence processing from the viewpoints of linguistics, psycholinguistics and computer science; * to synthesize findings about the mechanisms of Japanese sentence processing by researchers in these three fields in Japan and the United States; * to lay foundations for future interdisciplinary research in Japanese sentence processing, as well as international collaborations between researchers in Japan and the United States. The chapters in this volume have been written from the points of view of three different disciplines, with various immediate objectives -- from building usable speech understanding systems to investigating the nature of competence grammars for natural languages. All of the papers share the long term goal of understanding the nature of human language processing mechanisms. The book is concerned with two central issues -- the universality of language processing mechanisms, and the nature of the relation between the components of linguistic knowledge and language processing. This volume demonstrates that interdisciplinary research can be fruitful, and provides groundwork for further research in Japanese sentence processing.

Chapter 1

Japanese Sentence Processing: An Interdisciplinary Approach

Reiko Mazuka

Noriko Nagai

Duke University

Introduction

This volume is the result of the International Symposium on Japanese Sentence Processing held in October 1991 at Duke University. The symposium provided the first opportunity for researchers in three disciplines, from both Japan and the United States, to participate in a conference where they could discuss issues concerning Japanese syntactic processing. The goals of the symposium were three-fold:

1. To illuminate the mechanisms of Japanese sentence processing from the viewpoints of linguistics, psycholinguistics, and computer science.
2. To synthesize findings about the mechanisms of Japanese sentence processing by researchers in these three fields both in Japan and in the United States.
3. To lay foundations for future interdisciplinary research in Japanese sentence processing, as well as international collaborations between researchers in Japan and in the United States.

For the last two decades, mechanisms of natural language processing have been the focus of extensive investigation in psychology, linguistics, and computer science. In the field of psychology, the investigation of natural language processing is of crucial significance to our endeavors to understand general human cognitive mechanisms. For linguistics, in our attempt to investigate universal properties of grammatical knowledge, it is essential to investigate how grammatical knowledge interacts with performance; that is, how humans utilize their grammatical knowledge to comprehend and generate sentences. In computer science, models for natural language processing are mathematically laid out and implemented, and their validity is tested by actual simulation.

Research in natural language processing in the three fields just mentioned has only recently gained a forum in interdisciplinary journals and at conferences where researchers in the various fields can present their findings to an interdisciplinary audience. The success of such journals and conferences has shown that such interdisciplinary exchange is highly productive for progress in the field.

To date, the research in natural language processing has been mostly limited to English. Obviously, however, as the ultimate goal of the research lies in the understanding of human cognitive mechanisms, it is critical that cross-linguistic perspectives are represented. Japanese is an ideal language to investigate in contrast to English because it involves reverse forms of syntactic structures in exactly those dimensions that seem likely to be critical for sentence processing (i.e., head direction, presence of extensive empty categories, and presence of scrambling).

In theoretical linguistics, which recognizes the distinction between competence and performance, it is crucial to clarify the relationship between the domain of the two components. In the past, important works by Dryer (1980), J. D. Fodor (1978), and Kuno (1973) showed that performance influences competence. What has not been so clear is how performance influences competence. Recent works, such as that by Hawkins (1990), have just begun to address this issue. Another important issue that concerns the exact relation between competence and performance rests on the exact grammatical status of sentences. When a sentence is judged ungrammatical, we must correctly determine whether this judgment arises because the sentence violates grammatical conditions or if it is due to parsing difficulties. Unless we know exactly which grammatical phenomena fall into the domain of competence and which into the domain of performance, we can hardly begin to discuss the relation between the two components in any meaningful way, or even study competence effectively.

In computational linguistics, over the past few decades, numerous models have been put forward as models for human sentence processing (e.g., Frazier and Fodor’s Sausage Machine, 1978, Wood’s Augmented Transition Network model, 1970, Marcus’s Deterministic Parser, 1980). These models are assumed to characterize how humans process natural languages, and they are implicitly or explicitly assumed to be universal. Yet these models are based on detailed studies of English, leaving the study of other languages, especially non-Indo-European languages, largely neglected. When those “universal models” are applied to languages such as Japanese, which is head-final and allows extensive use of empty categories, they are found to make the unrealistic prediction that Japanese should be an extremely difficult language to process (see Mazuka & Lust, 1988, 1989). Thus, in order to capture the fact that the human natural language processing mechanism is capable of comprehending and generating both head-initial languages, such as English, and head-final languages, such as Japanese, with equal efficiency, it is essential to investigate how Japanese speakers process Japanese sentences and then integrate the Japanese results with the English data accumulated over the years.

In psycholinguistics, recent advances in methodologies for obtaining precise measures of human behavior during on-line comprehension and production of language, such as the eye-movement monitoring technique (cf. Rayner, 1983), the cross-modal priming technique (Swinney, 1979), and the on-line grammaticality judgment task (Kurtzman, Crawford, & Nychis-Florence, 1991), have allowed researchers to test finely tuned hypotheses with regard to how humans process natural language (see Mitchell, in press, for a review). The results of these psycholinguistic experiments can provide empirical validation as to whether a particular model for natural language processing is appropriate as a model of how humans comprehend or generate language. However, experimental investigation in sentence processing has concentrated almost entirely on English and a handful of other European languages, again neglecting the non-Indo-European languages, such as Japanese. Only very recently have researchers attempted to include Japanese in their inquiries.

Issues in Japanese Sentence Processing

Within these three fields, numerous theories of sentence processing have been put forth. There are two major issues that have concerned researchers of Japanese sentence processing. These issues extend beyond the domain of individual disciplines, and any attempt to resolve them requires collaborative endeavors across the disciplines.

The Universality of the Human Language Comprehension Mechanism

A central concern is the issue of whether the human sentence processing mechanism is in fact universal, which has been assumed either implicitly or explicitly by virtually all researchers in the field.

Everyone agrees that the human sentence processing mechanism must be universal to a certain degree, although the grammars of languages may vary systematically from one to another. The question is how to account for differences across languages within the framework of a universal theory of language processing.

Inoue and Fodor’s information-paced parser (this volume) provides a strong form of the universal hypothesis. According to their model, the human language comprehension mechanism is universal, and Japanese and English can be processed by exactly the same mechanism without any parameterization. Others have proposed that while the core of the parsing mechanism is universal, a few parameters must be incorporated into the basic parsing mechanism (Berwick & Fong, this volume; Mazuka, in press; Weinberg, this volume). However, the proposals differ in how the parameterization of the grammar interacts with the specific nature of the parsing mechanism. For example, according to Weinberg’s model, the parameterization of the parser is restricted to a minor component of the parser (i.e., whether IP is licensed directly by Infl[English] or indirectly licensed by a nominative marked NP [Japanese] is a parametric variation). In Mazuka (in press), parameterization concerns a more global aspect of parsing (i.e., whether clauses should be attached in relation to the topmost S on-line, or parsed to be the lowest S, until otherwise marked). Mazuka and Itoh’s (this volume) tentative attachment strategy implies that this strategy may not apply universally.

The Relation between Grammatical Knowledge and the Processing Mechanism

Another question that is critical in our endeavor concerns the exact nature of the relation between grammatical knowledge and the human language processing mechanism that utilizes such information. Because of its syntactic structure (as previously mentioned), Japanese provides an ideal language in which to investigate this issue.

Models differ in how they formalize the nature of our grammatical knowledge: Government and Binding theory (GB) based formalization (see, in this volume, Berwick & Fong; Gorrell, Inoue, & Fodor; Mazuka & Itoh; Nagai; Nakayama; Pritchett & Whitman; Sakamoto; Weinberg). Unification-theory-based formalization (Gunji; Hasida; Kameyama), or Functional-theory-based formalization (Kuno). Any model for human language processing must make an explicit assumption as to how such linguistic knowledge can be accessed by the parser. In this volume, the researchers’ main concerns center around three basic issues.

Types of Linguistic Representation. One issue concerns the validity of the linguistic representation in relation to the parsing mechanism. Gunji argues that if linguistic knowledge is represented in the formalization of Japanese Phrase Structure Grammar, which is in the class of context-free grammars, it is “computationally tractable,” thus more psychologically plausible than a GB-based grammar. Berwick and Fong, on the other hand, present a fully implemented computational model for a GB-based grammar, and show that a principles and parameters model of grammar is not computationally “intractable,” and can be investigated straightforwardly in its implications for parsing methods.

Types of Processing Components. A second issue concerns the choice of architecture for the parser. Given a particular type of linguistic representation, what types of parsing architecture can best account for the phenomena of Japanese sentence processing? For example, among the researchers using GB-based representation, Inoue and Fodor’s model proposes that incoming items are attached fully into a parse tree without delay in both Japanese and English. Pritchett (1992), on the other hand, has proposed a head-driven parser, where such an attachment decision is made at the grammatical head, which results in a delayed attachment decision in a head-final language such as Japanese. In Weinberg’s minimal commitment parser, following Marcus, Hindle, and Fleck’s (1983) D-theory parser, parsing decisions are made monotonically while individual parsing decisions are underspecified in such a way that inserting an additional node later in the parse will not require additional processing costs. Gorrell’s proposal, in which a parser builds a phrase-structure tree based on “dominance” and “precedence,” also assumes that parsing decisions are made monotonically.

Kameyama and Hasida each propose significantly different architectures to process Japanese, although both use Unification-theory-based representation. For example, Kameyama adopts a unification-based parser/interpreter called the Japanese Language Engine, which is a descendant of the Core Language Engine (Alshawi, Moore, & Moran, 1987), and Hasida uses a dynamic field of force to model the processing.

Competence/Performance. The dichotomy between grammatical knowledge and the parsing mechanism also has implications for generative theories that differentiate between the competence and performance. In such theories, the dichotomy between grammatical knowledge and parsing mechanisms roughly corresponds to the distinction between theories of competence and theories of performance. To clarify the relation between the two components, a number of questions must be resolved. In this volume, two specific issues relevant to this topic are discussed. The first issue addresses the exact relationship between the mental grammar and the parser. More specifically, we consider whether competence theory can predict the complexity of the human sentence mechanism (see chapter 3) and whether the parser has a transparent relation to the mental grammar (see chapters 11 and 12). Pritchett and Whitman (chapter 3) reassess the once-abandoned Derivational theory of complexity for predicting syntactic complexity. They argue that given the recent developments of linguistic theory (GB theory), in particular, trace theory and the addition of the logical form (LF) component, derivational history can correctly predict the syntactic complexity of a sentence.

Nakayama and Sakamoto (chapters 11 and 12) examine the psychological reality of linguistic theory via psycholinguistic experiments. Sakamoto, for example, considers the degree of transparency in the relationship between the mental grammar and performance theory, claiming that the most recent filler (MRF) strategy proposed by Frazier, Clifton, and Randall (1983) does not recognize the empty category as a possible antecedent (the lexical fillers only [LFO] hypothesis). Frazier et al. assume a nontransparent relationship between the parser and the mental grammar, the latter of which recognizes the empty category as a possible antecedent of PRO. This assumption is contrary to recent psycholinguistic research, which assumes a transparent relationship between the mental grammar and the parser. Sakamoto tests Japanese sentences against the MRF strategy and the LFO hypothesis and argues on the basis of the experimental results that either the MRF strategy or the LFO hypothesis is false for Japanese. Nakayama examines whether traces left by scrambling can be detected by probe recognition, as is possible with other types of traces (e.g., NP trace) and finds that traces left by scrambling cannot be detected using this technique.

The second issue addresses whether certain grammatical phenomena can be accounted for by an analysis based on performance. Nagai clarifies the grammatical status of “topics” which involve relative clauses. She argues that topicalization out of relative clauses, which has been analyzed in terms of linguistic theory (cf. Hasegawa, 1985) is more accurately analyzed from a performance perspective. She claims that all topicalization out of relative clauses is grammatical (i.e., licensed by grammatical principles). However, some topics result in unacceptable sentences due to parsing difficulties.

For theories such as Functional theory (cf. Kuno, 1987), the domain of linguistic investigation extends beyond the sentence level to include discourse. Such theories do not demarcate competence from performance. However, in this framework as well, the relationship between linguistic knowledge and the processing component is a central concern. Kuno’s chapter examines the appearance of null elements in parallel-sentence discourse and proposes a number of functional conditions. He suggests that these conditions are difficult to implement computationally.

Hasida extends the domain of this research even further and proposes that human language processing should be seen as a part of a general cognitive system, which he proposes to model using a dynamic field of force. His model is motivat...