Analysis of Japanese Compound Nouns using Collocational Information

Yosiyuki Kobayasi, Takenobu Tokunaga, Hozumi TanakaCOLING 1994

TL;DR

This paper tackles the challenge of automatically analyzing the structure of Japanese compound nouns, which lack explicit word boundaries. They propose a novel method combining word collocation statistics from a large corpus with thesaurus categories to determine the most likely structure, achieving around 80% accuracy. This approach is crucial for improving natural language processing systems that require a deep understanding of word relationships.

What Problem Does This Paper Address?

Natural language processing (NLP) systems often need to understand the meaning and structure of words. Compound nouns, formed by combining multiple words, are particularly challenging, especially in languages like Japanese where words are not separated by spaces. This lack of clear boundaries makes it difficult to even identify individual words within a compound, let alone understand how they relate to each other. Registering all possible compound nouns in a dictionary is impractical due to the infinite ways new compounds can be formed. Therefore, a method is needed to analyze the structure of these compounds from their constituent parts. Previous attempts often relied solely on statistical models, which could miss linguistic nuances, or made simplifying assumptions about word lengths.

Methods

The authors propose a method that leverages two key sources of information:

  • Word Collocational Information: They analyze how words appear together in a large corpus of Japanese text. Specifically, they focus on four-character sequences, which are common in Japanese compound nouns. By counting how often words co-occur, they can infer relationships between them.
  • Thesaurus Categories: To generalize beyond specific words and handle variations, they map words to broader categories from a Japanese thesaurus (Bunrui Goi Hyou - BGH). This allows them to analyze relationships between categories rather than just individual words.

The core of their method involves:

  1. Collecting Word Collocations: They extract four-character word sequences from a corpus.
  2. Assigning Thesaurus Categories: Each word in the collocations is assigned a category from the BGH thesaurus. They use a simple approach where if a word has multiple categories, they only consider cases where it has a single assigned category.
  3. Counting Category Collocations: They then count the co-occurrence frequencies of these thesaurus categories.
  4. Structure Analysis: For a given compound noun, they enumerate possible binary tree structures representing its composition. The preference for each structure is calculated recursively using the thesaurus category co-occurrence frequencies. They use two measures: simple probability (P(cat1, cat2)) and modified mutual information (MIS), which accounts for the position of words (left/right).

Datasets

  • Training/Collocation Corpus: A corpus of 160,000 four-character Japanese words, manually created from newspaper articles.
  • Thesaurus: Bunrui Goi Hyou (BGH), a standard machine-readable Japanese thesaurus, used at level 3 of its hierarchy.
  • Test Data: Manually created sets of compound nouns of four, five, and six kanji characters, extracted from newspaper editorials, columns, and encyclopedias. These test sets were not part of the training corpus.

Key Insights and Ideas

  • The Power of Generalization: Mapping words to thesaurus categories is crucial. It allows the system to handle variations in word choice and generalize patterns, making the analysis more robust than relying solely on specific word co-occurrences.
  • Modifier-Modifice Distance Matters: A key finding is that the distance between a modifier word and the word it modifies significantly influences the likelihood of a particular structure. Modifiers tend to modify closer words, a tendency that can be mathematically modeled and incorporated into the analysis.
  • Heuristics for Segmentation: Japanese word segmentation is inherently ambiguous. The paper uses a heuristic of "minimizing the number of content words" to prune the search space of possible segmentations, which is a common technique in Japanese morphological analysis.
  • Thesaurus Choice is Important: The authors note that their simple category assignment strategy works because BGH has relatively few words with multiple categories. For thesauruses with more ambiguity, more sophisticated methods would be needed.

Key Results

The proposed method achieved an accuracy of about 80% on compound nouns with an average length of 4.9 characters. Specifically, for the test data:

  • Four-kanzi character words: Accuracy of 96% (using CV1 measure).
  • Five-kanzi character words: Accuracy of 73% (using CV1 measure).
  • Six-kanzi character words: Accuracy of 62% (using CV1 measure).

The paper also shows that incorporating the modifier-modifice distance (using a redefined CV measure, CV') led to significant improvements, particularly for five and six-kanzi character collections. For instance, the accuracy for five-kanzi words increased to 79% and for six-kanzi words to 70% with the improved CV'.

Limitations and Future Work

  • Segmentation Ambiguity: While the method improves structure analysis, it still relies on accurate segmentation. The paper acknowledges that segmentation failure can occur due to missing words in dictionaries or limitations of the heuristics used.
  • Thesaurus Limitations: The current method's simplicity in handling words with multiple thesaurus categories might limit its performance with other thesauruses.
  • Grammatical vs. Semantic Heads: The assumption that the right branch of a subtree represents the category of the subtree is based on Japanese being a head-final language. However, this might not always capture the semantic head, especially with suffixes.
  • Future Work: The authors suggest incorporating other syntactic information (like affixes), using more advanced semantic information (like selectional restrictions), and applying the method to other syntactic structures such as dependency relations. They also plan to explore the use of semantic heads more deeply.

Summary updated

BibTeX
@inproceedings{kobayasi1994analysis,
  author = {Yosiyuki Kobayasi and Takenobu Tokunaga and Hozumi Tanaka},
  title = {{Analysis of Japanese Compound Nouns using Collocational Information}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1994},
}
Figure