Automatic Semantic Classification for Chinese Unknown Compound Nouns

Keh-Jiann Chen, Chao-jan ChenCOLING 2000

TL;DR

This paper introduces a novel similarity-based model for automatically classifying the semantic categories of unknown Chinese compound nouns. By representing morphological rules and leveraging example-based similarity, the system effectively disambiguates semantic meanings and evaluates the plausibility of new word formations, achieving promising accuracy rates.

What Problem Does This Paper Address?

Natural language processing (NLP) systems often struggle with "unknown words" – words not present in their lexicons. This is particularly challenging for Chinese due to its rich morphology, where single characters carry meaning and are often polysemous. Chinese compound nouns, formed by combining morphemes, are a frequent source of unknown words. Identifying their semantic categories is crucial for downstream tasks like information retrieval and machine translation. Existing methods face difficulties in handling the diverse morpho-syntactic structures, irregular semantic relationships between morphemes, and inherent ambiguities in Chinese.

Methods

The core of the proposed solution is an example-based similarity measure for semantic classification. The authors first establish morphological rules for Chinese compound nouns, which serve three purposes: defining the rules themselves, evaluating the "properness" of a compound, and disambiguating the semantic head of a compound.

The process involves:

  1. Morphological Analysis: Segmenting unknown compound nouns into their constituent morphemes using a longest-matching criterion.
  2. Head Determination: Identifying the head morpheme, typically the last morpheme in a compound.
  3. Semantic Category Prediction: For the head morpheme, the system leverages pre-defined morphological rules derived from existing compounds. These rules capture semantic relationships between modifiers and the head.
  4. Similarity Measurement: A similarity score is calculated between the input compound's modifier and the patterns in the learned morphological rules. This score reflects the semantic plausibility of the unknown compound.
  5. Ambiguity Resolution: The semantic category with the highest similarity score is assigned to the unknown compound. If the head morpheme isn't covered by the rules, it falls back to using semantic categories from a thesaurus (Chilin).

The paper also introduces two similarity measures: an "over-all-similarity" and a "maximal-similarity," which are combined for improved performance.

Datasets

  • Sinica Corpus: A 5 million-word Chinese corpus with word segmentation and part-of-speech tagging. This corpus was used to extract unknown words for testing.
  • CKIP Lexicon: An 80,000-entry Chinese lexicon with syntactic categories.
  • Chilin (Synonym Thesaurus): Contains about 70,000 words organized into 1428 semantic classes, used for semantic categorization.
  • Collocation Dictionary of Noun and Measure Words (CDNM): Lists collocating measure words for nouns, providing examples of compounds and their semantic categories.

The paper does not introduce a new benchmark dataset but uses existing resources for evaluation.

Key Insights and Ideas

  • Example-Based Learning for Semantics: The core idea is that the semantic plausibility of a new compound can be inferred from the similarity of its structure and morphemes to existing, well-formed compounds. This bypasses the need for explicit common-sense knowledge.
  • Morphological Rules as Semantic Blueprints: The learned morphological rules are not just syntactic but encode semantic relationships, acting as a knowledge base for disambiguation.
  • Head Morpheme as the Semantic Anchor: The model relies heavily on the head morpheme to determine the overall semantic category, a common characteristic of Chinese compound nouns.
  • Handling Ambiguity through Similarity: By calculating similarity scores across different potential semantic interpretations of a head morpheme, the system can select the most likely one.
  • Combining Similarity Measures: Using both "over-all" and "maximal" similarity measures, and weighting them, leads to better performance than either alone.

Key Results

The system was evaluated on unknown common nouns extracted from the Sinica corpus.

  • Correction Rate: Achieved a correction rate of 84% for the first hundred samples and 81% for the second hundred samples.
  • Error Analysis: Errors were categorized into:
  • Semantic selection error: Incorrectly choosing a semantic category for a polysemous head morpheme (most common error).
  • Non-compositional compounds: Compounds whose meaning isn't a direct sum of their parts.
  • Conjunctive structures: Compounds with a different structure than the assumed head-modifier.
  • Head-initial constructions: Compounds where the head is at the beginning (contrary to the model's assumption).
  • Coverage: The system covers about 71% of unknown noun compounds with its learned morphological rules.

Limitations and Future Work

  • Non-Compositional Compounds: The model struggles with compounds whose meanings are not semantically compositional.
  • Fine-Grained Classification: Some semantic classifications are too fine-grained, making clear distinctions difficult even for human annotators.
  • Data Sparsity: Insufficient sample data for some morphemes hinders the similarity-based model's effectiveness.
  • Unlisted Morphemes: Compounds with head morphemes not present in the system or the Chilin thesaurus remain unidentifiable.

Future work suggested includes:

  • Collecting more semantically non-compositional words and adding them to the lexicon.
  • Expanding the set of examples for each morpheme.
  • Performing deeper analysis of semantic relations between modifiers and heads.
  • Extending the model to handle compound verbs.
  • Integrating the system into a larger self-learning framework for linguistic and world knowledge acquisition.

Summary updated

BibTeX
@inproceedings{chen2000automatic,
  author = {Keh-Jiann Chen and Chao-jan Chen},
  title = {{Automatic Semantic Classification for Chinese Unknown Compound Nouns}},
  booktitle = {International Conference on Computational Linguistics},
  year = {2000},
}
Figure