Semantic Structure Analysis of Japanese Noun Phrases with Adnominal Particles

Akira Shimazu, Shozo Naito, Hirosato NomuraACL 1987 · DOI

Introduction

In Japanese linguistics, the adnominal particle no presents a significant challenge for natural language processing. Appearing in approximately 50% of all noun phrases in technical and journalistic texts, no serves as a semantically neutral connector in the "A no B" construction. Unlike English prepositions that specify relationships (e.g., "by," "on," "of"), no leaves the underlying semantic connection implicit. For instance, Furansujin no hanashi can be interpreted as either "a story about a Frenchman" or "a story told by a Frenchman." This inherent ambiguity necessitates a deep semantic analysis beyond simple syntactic parsing.

Methodology: Unification-Based Semantic Function Application

The proposed framework utilizes a unification-based semantic function application, drawing on principles from Categorial Unification Grammar (CUG). In this model, nouns are not merely static labels but can act as "functors" (functions) that take other nouns as arguments.

The system represents these relationships through complex feature structures—sets of attribute-value pairs including syntactic category (syncat), semantic category (semcat), word sense (sense), and predicate information (pred). The particle no acts as a functional operator that modifies the feature structure of the preceding noun (A), enabling it to unify with the head noun (B).

Semantic structures for "kooshijima no nuno" and "kooen no doozoo" illustrating feature-value unification.

A Taxonomy of Semantic Relations

Through an extensive study of over 10,000 examples from scientific journals and newspapers, the authors identified five primary semantic groups for the "A no B" construction. These groups are further subdivided into approximately 80 specific semantic relations.

  1. Predicate-Argument (B as Predicate): The head noun B acts as a predicate, and A is its argument (e.g., "his marriage").
  2. Case Role Restriction: B represents a location or time restricted by A (e.g., "front of a house").
  3. Attribute-Possession: B is an attribute of A (e.g., "weight of a stone").
  4. Predicate-Argument (A as Predicate): The modifying noun A acts as a predicate (e.g., "person who strolls").
  5. Attribute-Value: A provides an attribute value for B (e.g., "statue in a park").

To validate these classifications, the system attempts to paraphrase the noun phrases into explicit relative clauses, as shown in the table below.

The five primary semantic groups for "A no B" constructions with corresponding paraphrases.

Results and Contextual Disambiguation

A preliminary experiment involving 1,000 examples yielded satisfactory results, demonstrating the robustness of the unification-based approach. A key insight of this research is that many nouns are nominalized forms of verbs or adjectives, possessing "case frames" that dictate expected arguments.

Furthermore, the study highlights the role of context in resolving ambiguity. By unifying the semantic structure of a noun phrase with the event structure of preceding sentences, the system can distinguish between multiple valid interpretations, such as determining whether a "picture" is owned by an individual or created by them.

Conclusion and Future Directions

While the current method provides a granular map of Japanese noun-noun interactions, it remains dependent on highly detailed lexical entries for every noun. Future work aims to address the complexities of recursive "nested" structures (e.g., A no B no C) and to integrate broader contextual analysis, including analogy and association, to further refine the disambiguation process.

Summary updated

BibTeX
@inproceedings{shimazu1987semantic,
  author = {Akira Shimazu and Shozo Naito and Hirosato Nomura},
  title = {{Semantic Structure Analysis of Japanese Noun Phrases with Adnominal Particles}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {1987},
  doi = {10.3115/981175.981193},
}
Figure