An Automatic Clustering of Articles Using Dictionary Definitions

Fumiyo Fukumoto, Yoshimi SuzukiCOLING 1996

TL;DR

This paper introduces a novel method for automatically clustering articles by leveraging online dictionary definitions to disambiguate word senses and link semantically similar words. This approach significantly improves article classification, especially within specialized domains where word polysemy is a common challenge.

What Problem Does This Paper Address?

Automatically clustering articles, or grouping similar documents, is a fundamental task in information retrieval and natural language processing. While existing methods often rely on word frequencies to represent articles as vectors, they struggle with polysemy – the phenomenon where a single word can have multiple meanings. This is particularly problematic in specialized domains (like financial news) where certain words are frequently used with different senses. For instance, the word "security" can refer to physical safety or financial instruments, and failing to distinguish these senses can lead to misclassification. Furthermore, existing methods often treat words as independent units, overlooking the importance of phrasal lexicons (like "New York") which function as single semantic entities. This paper aims to address these limitations by developing a more nuanced approach to article clustering that accounts for word senses and semantic relationships.

Methods

The proposed method consists of two main stages:

  1. Word-Sense Disambiguation (WSD): This is the core innovation. The authors employ a WSD technique based on Niwa's method, which calculates the similarity between a sentence containing a polysemous word and its corresponding dictionary definitions. Specifically, they represent sentences as vectors using mutual information between words and then measure similarity using the inner product of normalized vectors. If a word has multiple senses defined in the dictionary, the sense that yields the highest similarity score is chosen. This process replaces polysemous words in the original articles with their disambiguated sense representations.
  2. Linking Nouns with Semantically Similar Nouns: To overcome the issue of reduced word frequencies after WSD and to capture semantic relationships, the authors introduce a linking mechanism. This involves:
  3. Calculating Mutual Information (Mu): Mu is calculated between noun pairs within the corpus to gauge their statistical association.
  4. Vector Representation of Nouns: Each noun is represented as a vector where the weight is derived from its Mu value with other nouns in the article.
  5. Measuring Dissimilarity: A dissimilarity measure (semantic deviation) is computed between noun vectors.
  6. Clustering: Nouns are clustered based on their semantic deviation values using a non-overlapping, group average method.
  7. Linking: Nouns that are semantically similar (belonging to the same cluster) are linked and represented by a single representative word. This effectively creates a form of phrasal lexicon.

Finally, articles are represented as vectors based on the frequencies of these disambiguated and linked words, and then clustered using a similarity measure.

Datasets

  • Training Corpus: 1988 and 1989 WSJ (Wall Street Journal) corpus from ACL/DCI CD-ROM, containing approximately 280,000 part-of-speech tagged sentences.
  • Test Data: 49 randomly selected articles from the WSJ corpus, each with a different topic tag.
  • Dictionary: Collins English Dictionary from ACL/DCI CD-ROM.

Key Insights and Ideas

  • Dictionary Definitions as a Powerful Tool for WSD: The paper demonstrates that leveraging dictionary definitions, not just for word meanings but for their contextual similarity to sentences, is a highly effective way to disambiguate word senses.
  • Bridging the Gap Between WSD and Clustering: The proposed method seamlessly integrates WSD into the article clustering pipeline, directly addressing the problem of polysemy that plagues simpler frequency-based methods.
  • Semantic Linking Enhances Representation: By linking semantically similar nouns, the method creates a more robust representation of articles, capturing phrasal meaning and mitigating the loss of information due to reduced word frequencies after WSD.
  • Effectiveness in Specialized Domains: The approach shows particular promise for classifying articles within restricted domains, where word sense ambiguity is more pronounced.

Key Results

The paper reports on four experimental setups: 'Freq' (frequency-based, no WSD/linking), 'Dis' (WSD only), 'Link' (linking only), and 'Method' (proposed WSD + linking).

  • Overall Performance: The proposed 'Method' achieved the highest clustering accuracy, correctly classifying 31 out of 40 sets (77.5%). This significantly outperforms 'Freq' (52.5%), 'Link' (57.5%), and 'Dis' (60.0%).
  • Impact of WSD: The 'Dis' experiment showed a notable improvement over 'Freq', correctly classifying 24 sets compared to 21. This highlights the benefit of disambiguating word senses.
  • Impact of Linking: The 'Link' experiment also showed improvement over 'Freq', correctly classifying 23 sets. This suggests that capturing semantic relationships between words is beneficial.
  • Qualitative Findings: Figures in the paper illustrate that the 'Method' correctly groups articles into categories like 'food restaurant' and 'market news', whereas simpler methods misclassify them due to polysemous words. For example, the word "security" was correctly disambiguated in different contexts, leading to better classification.

Limitations and Future Work

  • Limited Definitions: The authors acknowledge that they used only the top 5 definitions for each noun from the dictionary. Some words might have meanings not covered by these top definitions, potentially hindering higher clustering accuracy.
  • Alternative Weighting Policies: The paper suggests exploring alternative weighting policies, such as weighted inverse document frequency (widf), which has shown advantages over inverse document frequency (idf) in text categorization.
  • Using All Definitions: A potential improvement is to utilize all available definitions for words in the dictionary.
  • Further Experiments: The authors plan to conduct further experiments to address the remaining issues and apply the method to practical use.

Summary updated

BibTeX
@inproceedings{fukumoto1996automatic,
  author = {Fumiyo Fukumoto and Yoshimi Suzuki},
  title = {{An Automatic Clustering of Articles Using Dictionary Definitions}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1996},
}
Figure