A System of Verbal Semantic Attributes Focused on the Syntactic Correspondence between Japanese and English

Hiromi Nakaiwa, Akio Yokoo, Satoru IkeharaCOLING 1994

TL;DR

This paper introduces a novel system of 97 verbal semantic attributes designed to bridge the gap between Japanese and English verb meanings and their usage. By focusing on both a verb's dynamic characteristics and its relationship with grammatical cases, this system effectively disambiguates verb meanings in a large-scale Japanese-English transfer pattern dictionary, paving the way for more accurate machine translation.

What Problem Does This Paper Address?

Machine translation systems, while increasingly practical, still struggle with the nuances of language, particularly with phenomena like ellipsis and anaphoric references that require deep contextual understanding. A major hurdle is the sheer volume of knowledge needed to accurately determine word meanings and their relationships within sentences. This paper tackles the challenge of verb meaning disambiguation, which is crucial for understanding sentence flow and achieving high-quality translations. Specifically, it addresses the difficulty of mapping Japanese verbs to their English equivalents, especially when dealing with the polysemy (multiple meanings) of verbs and the significant syntactic differences between the two languages. Existing verb classification methods often focus solely on word meaning without adequately considering how verbs are used in context, limiting their utility for natural language processing.

Methods

The core of this paper is the proposal of a system of 97 verbal semantic attributes. These attributes are derived from two primary classification criteria:

  1. Dynamic Characteristics of Verbs: This categorizes verbs based on their inherent meaning and the type of action or state they represent (e.g., possession, production, physical transfer, mental action).
  2. Relationship of Verbs to Cases: This classifies verbs based on the roles they assign to their grammatical case elements (e.g., subject, direct object, indirect object).

These attributes are then applied to a Japanese-English transfer pattern dictionary containing 15,000 pairs of Japanese valence patterns and their corresponding English syntactic structures. The system aims to disambiguate the meanings of these verb-pattern pairs by assigning specific semantic attributes. The authors also highlight the importance of pairing Japanese and English patterns to capture common meanings and usage, thereby reducing conceptual ambiguities.

Datasets

The primary dataset used in this research is a Japanese-English transfer pattern dictionary comprising approximately 15,000 pairs. These pairs consist of Japanese valence patterns and their equivalent English syntactic structures. The paper does not introduce a new benchmark dataset but rather utilizes an existing one for evaluation.

Key Insights and Ideas

  • Verb meaning is deeply tied to usage: The paper strongly emphasizes that understanding a verb's meaning requires considering its syntactic context and how it interacts with other sentence elements (cases).
  • A unified attribute system is powerful: By developing a set of 97 verbal semantic attributes that capture both dynamic characteristics and case relationships, the authors demonstrate a robust way to disambiguate verb meanings across languages.
  • Pattern pairing is key for cross-lingual mapping: Matching Japanese and English patterns that share common meanings and usage is crucial for overcoming syntactic differences and accurately translating verb semantics.
  • Contextual analysis is essential for zero pronouns: The paper shows that verbal semantic attributes are highly effective in resolving anaphoric references for Japanese zero pronouns, a common challenge in NLP.
  • Semantic attributes aid in handling ellipsis: The system can help supplement missing case elements in elliptical sentences by leveraging the semantic constraints provided by verb attributes.

Key Results

  • The proposed system of 97 verbal semantic attributes was applied to a 15,000-pair Japanese-English transfer pattern dictionary.
  • Approximately 90% of these patterns could be described by a single attribute value, indicating the system's efficiency and descriptive power.
  • The study found that 73.4% of verbs in the dictionary had multiple patterns, and for 70.1% of these, different kinds of verbal semantic attributes were added to the patterns, highlighting the necessity of such a nuanced system.
  • The top 10 most frequent verbal semantic attributes cover a significant portion of the patterns, with "ATTRIBUTE" (27.0%) and "BODILY ACTION" (12.7%) being the most prevalent.
  • In an application for anaphora resolution of Japanese zero pronouns, the system achieved a success rate of 95% in a blind test, outperforming a previously proposed method.

Limitations and Future Work

  • Need for finer granularity: The authors acknowledge that while the system is effective, some attributes (like "BODILY ACTION" and "ATTRIBUTE TRANSFER") still present challenges in resolving semantic ambiguities, suggesting a need for further subdivision of these attributes in the future.
  • Expansion to other NLP tasks: The paper suggests that verbal semantic attributes can be applied to other context processing problems, including estimating tense relationships between events and sentence abridgment.
  • Further refinement of standards: The authors are continuing to work on refining the details and tightening the standards of the system, as well as exploring more extensive applications.

Summary updated

BibTeX
@inproceedings{nakaiwa1994system,
  author = {Hiromi Nakaiwa and Akio Yokoo and Satoru Ikehara},
  title = {{A System of Verbal Semantic Attributes Focused on the Syntactic Correspondence between Japanese and English}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1994},
}
Figure