A Statistical Approach to the Processing of Metonymy

Masao Utiyama, Masaki Murata, Hitoshi IsaharaCOLING 2000

TL;DR

This paper presents a statistical method for interpreting metonymy in Japanese, a common linguistic phenomenon where one thing stands for another related thing. By leveraging corpus statistics, the approach ranks potential interpretations of metonymic phrases, achieving a promising 53% accuracy on a diverse set of 75 metonymies. This work offers a data-driven alternative to knowledge-intensive methods for handling metonymy in natural language processing.

What Problem Does This Paper Address?

Metonymy, where a word or phrase stands for something closely related (e.g., "Shakespeare" for "the works of Shakespeare"), is pervasive in natural language. Accurately processing metonymy is crucial for applications like machine translation, as a metonymy acceptable in one language might be nonsensical in another. Previous approaches often relied on hand-crafted ontologies or semantic networks, which are difficult to build and can limit applicability to specific domains. This paper addresses the challenge of developing a more general and data-driven method for interpreting metonymy, specifically focusing on Japanese.

Methods

The proposed method involves two main steps: recognition and interpretation. While the paper acknowledges the importance of metonymy recognition (identifying metonymic expressions), its primary focus is on interpretation. For a given metonymic phrase of the form "Noun A Case-Marker R Predicate V," the system first extracts candidate nouns that can be syntactically related to the explicit term A from a large corpus. These candidates are then ranked using a statistical measure, $M(B|A, R, V)$, which estimates the appropriateness of a noun B as an interpretation. This measure is derived from probabilities calculated using corpus frequencies, specifically considering the "A no B" (B of A) and "A near B" (nouns within a certain proximity) syntactic relations. The method leverages conditional probabilities and frequency counts from the corpus to quantify the likelihood of a candidate noun being the intended implicit term.

Datasets

  • Corpus: A large corpus of seven years of the Mainichi Newspaper (1991-1997), consisting of approximately 153 million words. This corpus was morphologically analyzed using ChaSen version 2.0b6.
  • Metonymy Set: Seventy-five metonymies were used for evaluation. Sixty-two were collected from Japanese literature on cognitive linguistics and psycholinguistics, chosen for their diverse types. The remaining 13 were direct translations of English metonymies from prior research.
  • Semantic Class Thesaurus: The Bunrui Goi-Hyô (The National Language Research Institute, 1996) was used to categorize nouns into a six-layered hierarchy, containing over 55,000 nouns. This was used to explore the effectiveness of semantic class information.

Key Insights and Ideas

  • Corpus Statistics as a Powerful Tool: The paper demonstrates that corpus statistics can effectively capture the relationships needed to interpret metonymy, offering a viable alternative to labor-intensive knowledge engineering.
  • "A no B" Relation is Key: The "A no B" syntactic relation in Japanese proved particularly effective for filtering out loosely related words and identifying strong semantic candidates for metonymic interpretation.
  • Combining Relations is Beneficial: Using both "A no B" and "A near B" relations for candidate extraction yielded better accuracy than using either relation in isolation, suggesting that a broader syntactic context can improve results.
  • Asymmetrical Nature of Metonymy: The statistical formulation implicitly acknowledges that the relationship between the explicit term and the implicit term is often asymmetrical, with the implicit term having a tighter connection to the explicit term than to the predicate.

Key Results

  • The proposed statistical method correctly interpreted 53 out of 75 metonymies in Japanese.
  • Using both "A no B" and "A near B" relations achieved an accuracy of 71% (53/75), with a 95% confidence interval of 0.61-0.81.
  • Using only the "A no B" relation resulted in 50 correct interpretations, while using only "A near B" yielded 43 correct interpretations.
  • For a subset of 13 directly translated metonymies, the method achieved 10 correct interpretations. When restricted to metonymies acceptable in Japanese, the accuracy rose to 90% (9/10), suggesting these were more typical and easier examples.

Limitations and Future Work

  • Recognition vs. Interpretation: The paper focuses on interpretation and does not deeply explore metonymy recognition, leaving the question of whether statistical methods can match selection-restriction violation methods for recognition as an open research area.
  • Semantic Relation Ambiguity: The paper notes that while "A no B" phrases are extracted, their precise semantic relations are not explicitly determined, requiring further semantic analysis for deeper understanding.
  • Effectiveness of Semantic Classes: The use of semantic classes for interpretation showed no significant improvement in the experiment, though the authors suggest further research is needed to fully judge their effectiveness, especially for cases where co-occurrence probabilities are zero.
  • Smoothing Techniques: The authors acknowledge that more sophisticated probability smoothing techniques could be beneficial but are beyond the scope of this paper.
  • Generalizability: While the results are promising, the authors suggest that the degree of accuracy achieved is likely to be repeatable for other general sets of metonymies, but further testing on broader datasets would be valuable.

Summary updated

BibTeX
@inproceedings{utiyama2000statistical,
  author = {Masao Utiyama and Masaki Murata and Hitoshi Isahara},
  title = {{A Statistical Approach to the Processing of Metonymy}},
  booktitle = {International Conference on Computational Linguistics},
  year = {2000},
}

Loading related content...

Figure