Combining a Chinese Thesaurus with a Chinese Dictionary

Donghong Ji, Junping Gong, Changning HuangACL 1998

TL;DR

This paper presents an automatic method to link Chinese thesaurus entries with Chinese dictionary senses, creating a richer semantic resource. By leveraging a "similar word strategy" based on definition analysis, they enable a thesaurus to specify sense similarity and a dictionary to gain semantic classification.

What Problem Does This Paper Address?

Creating comprehensive semantic resources like WordNet for English is labor-intensive. For many other languages, including Chinese, such integrated resources are lacking. Specifically, Chinese thesauri (like 《同义词词林》) group words by semantic categories but don't specify which sense of a polysemous word belongs to a category. Conversely, Chinese dictionaries (like 《现代汉语词典》) define senses but lack semantic classification. This paper tackles the challenge of automatically bridging this gap by aligning thesaurus entries with dictionary senses, thereby enriching both resources.

Methods

The core of the proposed method is a "similar word strategy" that extends the traditional "common word strategy." Instead of relying on shared words in definitions, it assumes that definitions of similar senses will contain similar words.

  1. Salience Calculation: For each thesaurus category (X), they first identify "salient codes" (semantic categories) that are most relevant to its member words. This is done by calculating the "definition salience" of a semantic code with respect to a category.
  2. Context Vectors: They build "context vectors" for categories based on the distribution of their member words in a large corpus, using mutual information between words and semantic codes.
  3. Distance Metrics: Distances are defined between categories and between semantic codes and senses. A key metric, dis(X, s), measures the distance between a category X and a dictionary sense s, weighted by the salience of semantic codes within X.
  4. Tagging: A sense s is tagged to a word w if dis(X, s) is below a learned threshold τ*, where X is a category w belongs to.

Datasets

  • Chinese Thesaurus: 《同义词词林》 (TongYiCi CiLin) - used for semantic categories.
  • Chinese Dictionary: 《现代汉语词典》 (XianDai HanYu CiDian) - used for word senses and definitions.
  • Corpus: An 80-million-character Chinese corpus was used for calculating word and code distributions.

The paper also implicitly uses a subset of univocal words (words with a single sense in the dictionary and a single code in the thesaurus) for parameter estimation and evaluation.

Key Insights and Ideas

  • Beyond Common Words: The "similar word strategy" is crucial. It acknowledges that direct word overlap might be rare in definitions of related senses, but semantic similarity can still be captured by similar vocabulary.
  • Definition Context is Special: The authors argue that definitions provide richer semantic information (synonyms, hyponyms) than general contexts, making direct application of general word sense disambiguation models less effective.
  • Bridging the Gap: The method effectively links abstract semantic categories from a thesaurus to concrete word senses in a dictionary, creating a more structured and informative semantic resource.
  • Indirect Evaluation: Evaluating the tagging procedure directly is difficult. The paper proposes an indirect evaluation by creating "artificial" entries and assessing how well the system tags them, using "None" as a correct tag when no appropriate sense exists.

Key Results

  • Parameter Estimation: The threshold τ* for dis(X, s) was estimated to be 0.384. This resulted in tagging 84.68% of univocal words with their unique senses.
  • Overall Tagging: The system processed 29,679 shared words, with 35,193 thesaurus entries and 36,426 dictionary senses.
  • Accuracy and Loss:
    • Under different constraints for creating new entries, average accuracy rates ranged from 88.39% to 95.26%, with corresponding loss rates from 11.61% to 4.74%.
  • For entries with varying numbers of tags (0, 1, 2), average accuracy rates were 94.6%, 90.1%, and 87.6%, respectively.

Limitations and Future Work

  • No Syntactic Parsing: The method does not perform syntactic parsing on definitions, which the authors acknowledge could be a limitation. Distinguishing definitions based on syntactic structures (e.g., sub-obj vs. "of" structures) requires more advanced analysis.
  • Incomplete Resources: The electronic versions of the resources used contained only parts of the original data.
  • Noisy Semantic Codes: Noisy semantic codes in definitions can affect tagging accuracy, though the probability is considered lower when multiple salient codes exist.
  • Future Work: The authors suggest that the method is appropriate for the general problem of identifying senses that account for similarity in a set of words. They also note that the soundness of the results improves with more words in a category and greater similarity among them. The paper also hints at future work in exploring syntactic structures for better disambiguation.

Summary updated

BibTeX
@inproceedings{ji1998combining,
  author = {Donghong Ji and Junping Gong and Changning Huang},
  title = {{Combining a Chinese Thesaurus with a Chinese Dictionary}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {1998},
}
Figure