Acquiring Sense Tagged Examples using Relevance Feedback

Proceedings of the 22nd International Conference on Computational Linguistics (Coling 2008), pages 809–816 Manchester, August 2008 Acquiring Sense Tagged Examples using Relevance Feedback Mark Stevenson, Yikun Guo and Robert Gaizauskas Department of Computer Science University of Sheffield Regent Court, 211 Portobello Sheffield, S1 4DP United Kingdom inital.surname@dcs.shef.ac.uk Abstract Supervised approaches to Word Sense Dis- ambiguation (WSD) have been shown to outperform other approaches but are ham- pered by reliance on labeled training ex- amples (the data acquisition bottleneck). This paper presents a novel approach to the automatic acquisition of labeled examples for WSD which makes use of the Informa- tion Retrieval technique of relevance feed- back. This semi-supervised method gener- ates additional labeled examples based on existing annotated data. Our approach is applied to a set of ambiguous terms from biomedical journal articles and found to significantly improve the performance of a state-of-the-art WSD system. 1 Introduction The resolution of lexical ambiguities has long been considered an important part of the process of understanding natural language. Supervised ap- proaches to Word Sense Disambiguation (WSD) have been shown to perform better than unsuper- vised ones (Agirre and Edmonds, 2007) but require examples of ambiguous words used in context an- notated with the appropriate sense (labeled exam- ples). However these often prove difficult to obtain since manual sense annotation of text is a complex and time consuming process. In fact, Ng (1997) estimated that 16 person years of manual effort would be required to create enough labeled exam- ples to train a wide-coverage WSD system. This c⃝ 2008. Licensed under the Creative Commons Attribution-Noncommercial-Share Alike 3.0 Unported li- cense (http://creativecommons.org/licenses/by-nc-sa/3.0/). Some rights reserved. limitation is commonly referred to as the data ac- quisition bottleneck. It is particularly acute in spe- cific domains, such as biomedicine, where terms may have technical usages which only domain ex- perts are likely to be aware of. For example, pos- sible meanings of the term “ganglion” in UMLS (Humphreys et al., 1998) include ‘neural structure’ or ‘benign mucinous tumour’, although only the first meaning is listed in WordNet. These domain- specific semantic distinctions make manual sense annotation all the more difficult. One approach to the data acquisition bottleneck is to generate labeled training examples automat- ically. Others, such as Leacock et al. (1998) and Agirre and Mart´ınez (2004b), used information from WordNet to construct queries which were used to retrieve training examples. This paper presents a novel approach to this problem. Rele- vance feedback, a technique used in Information Retrieval (IR) to improve search results, is adapted to identify further examples for each sense of am- biguous terms. These examples are then used to train a semi-supervised WSD system either by combining them with existing annotated data or using them alone. The approach is applied to a set of ambiguous terms in biomedical texts, a domain for which existing resources containing labeled ex- amples, such as the NLM-WSD data set (Weeber et al., 2001), are limited. The next section outlines previous techniques which have been used to …

Content truncated to 500 words