Word Domain Disambiguation via Word Sense Disambiguation

Antonio Sanfilippo, Stephen Tratz, Michelle GregoryNAACL 2006

TL;DR

This paper proposes a novel approach to Word Domain Disambiguation (WDD) by leveraging existing Word Sense Disambiguation (WSD) techniques. Instead of developing specialized WDD algorithms, the authors demonstrate that by accurately identifying a word's sense, its corresponding subject domain can be effectively inferred, achieving competitive results.

What Problem Does This Paper Address?

In natural language processing, understanding the specific meaning of a word in context is crucial. This is often tackled through Word Sense Disambiguation (WSD), where the goal is to identify the correct sense of a word from a lexicon like WordNet. Related to this is Word Domain Disambiguation (WDD), which aims to assign a subject domain (e.g., "finance," "medicine") to a word based on its usage. While WDD has been shown to improve WSD performance, dedicated WDD algorithms have received less attention. The challenge lies in the fact that a single word can belong to multiple domains depending on its context, and existing WDD approaches often require specialized algorithms. This paper questions whether WDD can be solved by simply re-framing it as a WSD problem.

Methods

The core of this paper's methodology is to use an existing Word Sense Disambiguation (WSD) algorithm as a stepping stone to achieve Word Domain Disambiguation (WDD). The process involves two main steps:

  1. Word Sense Disambiguation: An input text is processed by a WSD algorithm to assign a specific WordNet sense to each word. The authors utilize a supervised classification approach based on the maximum entropy algorithm, trained on the SemCor corpus. This algorithm considers contextual, syntactic, and semantic information to predict the most likely sense for a word.
  2. Word Domain Mapping: Once a word's sense is identified, the WordNet Domains lexicon is used to map this sense to its associated subject domain(s). WordNet Domains is an extension of WordNet that annotates synonym sets with subject domain labels.

This approach effectively transforms the WDD task into a WSD task, followed by a straightforward lookup in the WordNet Domains lexicon.

Datasets

  • SemCor: Used for training the WSD algorithm and as a gold standard for evaluating WDD results. The authors extended the training data by including examples from the Open Mind Word Expert corpus.
  • Senseval3: Used for evaluating the WDD approach, particularly for verbs and adjectives. This corpus was stripped of its sense annotations and processed using the WSD algorithm.

Key Insights and Ideas

  • WDD as a WSD problem: The most significant insight is that Word Domain Disambiguation can be effectively addressed by leveraging Word Sense Disambiguation, eliminating the need for specialized WDD algorithms.
  • Leveraging existing WSD power: By using a strong WSD system, the paper demonstrates that accurate word sense identification directly translates to accurate word domain assignment.
  • Extending WDD to verbs and adjectives: Unlike previous WDD work that primarily focused on nouns, this approach successfully extends WDD to verbs and adjectives.

Key Results

The paper reports strong results for their WDD via WSD approach:

  • SemCor: Achieved high accuracy across nouns, verbs, and adjectives, significantly outperforming a baseline that assigns the first sense of a word as its domain. For example, accuracy for nouns was 0.874, verbs 0.933, and adjectives 0.942.
  • Senseval3: Also showed competitive results, with accuracies of 0.797 for nouns, 0.908 for verbs, and 0.848 for adjectives. While the comparison to the baseline was less conclusive on Senseval3, the authors attribute this to the smaller size of the corpus.
  • The authors claim their approach compares favorably with previous WDD work, citing results from Bagnini and Strapparava (2000) and Suarez & Palomar (2002).

Limitations and Future Work

  • Adverbs: The authors explicitly state that they have not yet worked on adverbs but expect comparable results.
  • DSO Corpus: A meaningful comparison with Suarez and Palomar (2002) is hindered by the use of different datasets. The authors are working on repeating their study with the DSO corpus to enable direct comparison.
  • Bootstrapping Cycle: The paper suggests a promising future direction: a bootstrapping cycle where WDD results are fed back to improve the WSD model, which in turn could lead to even better WDD results. This iterative improvement is a key area for future exploration.

Summary updated

BibTeX
@inproceedings{sanfilippo2006word,
  author = {Antonio Sanfilippo and Stephen Tratz and Michelle Gregory},
  title = {{Word Domain Disambiguation via Word Sense Disambiguation}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2006},
}
Figure