Abstract
In this paper we concentrate on the resolution of the lexical ambiguity that arises when a given word has several different meanings. This specific task is commonly referred to as word sense disambiguation (WSD). The task of WSD consists of assigning the correct sense to words using an electronic dictionary as the source of word definitions. We present two WSD methods based on two main methodological approaches in this research area: a knowledge-based method and a corpus-based method. Our hypothesis is that word-sense disambiguation requires several knowledge sources in order to solve the semantic ambiguity of the words. These sources can be of different kinds--- for example, syntagmatic, paradigmatic or statistical information. Our approach combines various sources of knowledge, through combinations of the two WSD methods mentioned above. Mainly, the paper concentrates on how to combine these methods and sources of information in order to achieve good results in the disambiguation. Finally, this paper presents a comprehensive study and experimental work on evaluation of the methods and their combinations.
Introduction
Word-Sense Disambiguation (WSD) is the challenge of determining which "sense" of a word is being used in a specific context. Historically, researchers have used two main strategies: knowledge-based (KB) methods that use external resources like WordNet, and corpus-based (CB) methods that use supervised machine learning. While KB methods are scalable, they often struggle with precision; conversely, CB methods are accurate but suffer from the "knowledge acquisition bottleneck," requiring massive amounts of manually annotated text.
This paper explores how to break this bottleneck by combining these traditionally separate approaches. By integrating these techniques—using one to pre-process data for the other or combining them in a voting scheme—the authors demonstrate that hybrid systems significantly outperform standalone models.
Methodology
The authors present two primary systems and experiment with three ways to combine them.
Specification Marks Method (SM)
This knowledge-based method utilizes the WordNet hierarchy. It analyzes the nouns in a sentence to find the most specific concept in the taxonomy that "subsumes" or connects them, effectively identifying the most logical semantic cluster.
Maximum Entropy (ME)
This corpus-based supervised learning model represents context as feature vectors (word collocations, parts of speech, etc.). A key innovation is the use of "collapsed features," which group similar linguistic data to reduce the "sparsity" problem where the model sees a feature too rarely to learn from it.
Hybrid Approaches
- CB helping KB: The ME model disambiguates nouns in the context of a target word, providing the SM model with a "cleaner" set of starting points.
- KB helping CB: The SM model's domain information (e.g., identifying a text as "Biology") is fed into the ME model as an additional feature.
- Voting System (vME+SM): A majority voting scheme where multiple ME classifiers and the SM model all vote on the correct sense.
Experimental Results and Insights
The study utilizes standard benchmarks like SemCor and SENSEVAL-2. A striking insight is that these systems fail in different ways. Even though the ME model was generally more accurate, there were many cases (about 15% in the Spanish test) that the "weaker" SM model solved correctly while the ME model failed.
In the Spanish SENSEVAL-2 task, the hybrid vME+SM system achieved an accuracy of 68.4%, placing it second overall among international competitors and outperforming the standalone ME model (65.8%). Pre-processing context with ME improved the SM model's precision by about 3.6%, while adding domain information to the ME model provided a statistically significant boost for words with well-differentiated domains.
Conclusion
The hybrid approach proves to be a robust strategy for multi-lingual NLP. While the current SM method is primarily noun-centric due to its reliance on the WordNet "IS-A" hierarchy, future work aims to expand this to verbs and adjectives. The authors suggest that tighter integration of knowledge-based signals and the use of unsupervised corpus-based methods could further reduce human intervention in the disambiguation process.
Summary updated