<?xml version="1.0" standalone="yes"?>
<Paper uid="J98-1005">
  <Title>Disambiguating Highly Ambiguous Words</Title>
  <Section position="2" start_page="0" end_page="126" type="abstr">
    <SectionTitle>
1. Introduction
</SectionTitle>
    <Paragraph position="0"> An information retrieval system returns documents presumed to be of interest to the user in response to a query. While there are a variety of different ways the retrieval can be accomplished, most systems treat the query as a pattern to be matched by documents. Unfortunately, the effectiveness of these word-matching systems is depressed by both homographs and synonyms. Homographs depress the accuracy of the retrieval systems by making texts about two different concepts appear to match. Synonyms impair the system's ability to find all matching documents, since different words mask conceptual matches. While polysemy is the immediate cause of the first problem, it indirectly contributes to the second problem as well by preventing the effective use of thesauri. These considerations motivate our desire for a highly accurate word sense disambiguator.</Paragraph>
    <Paragraph position="1"> Our experimental results show that the disambiguator described in this paper is quite accurate. The disambiguator is a particular formulation of feed-forward neural networks (Rumelhart, Hinton, and Williams 1986) that separately extract topical and local contexts of a target word from a set of sample sentences that are tagged with the correct sense of the target. The neural networks responsible for topical and local disambiguation are then combined to form a single, &amp;quot;contextual&amp;quot; representation (Miller and Charles 1991). Further experiments show that the accuracy of the contextual disambiguator can be improved if the disambiguator is allowed to label some examples as  * Siemens Corporate Research, 755 College Road East, Princeton, NJ 08540 t Currently at: National Institute of Standards and Technology, Building 225, Room A-216, Gaithersburg, MD, 20899 (~) 1998 Association for Computational Linguistics  Computational Linguistics Volume 24, Number 1 unknown. Since the accumulation of sufficient tagged samples is expensive and timeconsuming, we finish by describing an extension of our algorithm through which its accuracy can be enhanced by using inexpensive untagged examples.</Paragraph>
    <Paragraph position="2"> Our long-term goal is to be able to incorporate such a contextual disambiguation system within a taxonomy such as WordNet (Miller 1990) and thereby to use it for resolving query word senses at retrieval run-time. To accomplish this goal, the disambiguator must be able to construct contextual representations that accurately distinguish among the highly ambiguous words found in general-purpose lexicons as well as build representations that are efficient to use at query run-time. The system described in this paper represents a significant step towards that goal.</Paragraph>
    <Paragraph position="3"> 2. Effects of Polysemy on Retrieval Performance The effectiveness of information retrieval systems is usually measured in terms of precision, the percentage of retrieved documents that are relevant, and recall, the percentage of relevant documents that are retrieved. As mentioned above, in principle, the direct effect of polysemy on word-matching systems is to decrease precision (e.g., queries about financial banks retrieve documents about rivers). The impact this direct effect has in practice is less clear. Schtitze and Pedersen (1995) found noticeable improvement in precision using sense-based (as opposed to word-based) retrieval. On the other hand, Krovetz and Croft (1992) concluded that polysemy hurt retrieval only if the searcher needed very high recall or was using very short (one or two word) queries. Sanderson (1994) found that resolving senses could degrade retrieval performance unless the disambiguation procedure was very accurate, although he worked with large, rich queries. Other techniques also address the polysemy problem without requiring explicit disambiguation. One such technique is local-global matching (Salton and Buckley 1991), where the similarity of a document with a query depends not only on the words occurring in the entire document but also on the existence of smaller lexical units, such as sentences, that exhibit particularly close matches with the query. These techniques implicitly accommodate ambiguity: by computing similarity measures based on word co-occurrence, the systems find instances of words used in the same contexts and thus words that are used in the same sense.</Paragraph>
    <Paragraph position="4"> Polysemy has a second, indirect effect, however, in that it hampers the successful application of thesauri. Much as using the same word in different senses can depress precision by causing false matches, using different words to express the same sense (i.e., synonyms) depresses recall by causing true conceptual matches to be missed.</Paragraph>
    <Paragraph position="5"> One way to mitigate the effects of synonyms is to use lexical aids to expand a text (usually the query) by words that are closely related to words in the original text.</Paragraph>
    <Paragraph position="6"> This procedure has met with some success in experiments on small, single-domain collections. For example, Salton and Lesk (1971) found that expansion by synonyms only improved performance, and Wang, Vendendorpe, and Evens (1985) found that a variety of lexical-semantic relations improved retrieval performance. However, it is difficult to obtain similar improvements in heterogeneous collections where the lexical aids necessarily contain multiple senses of words (Voorhees and Hou 1993; Voorhees 1994a, 1994b).</Paragraph>
    <Paragraph position="7"> Selecting the correct sense of a word to be expanded is an essential first step when expanding queries by lexically related words. Experiments run on the diverse TREC test collection (Harman 1993), and using the WordNet lexical system as a source of related words, demonstrate that expanding queries by the set of words such that each word in the set is related to some (any) sense of the query word consistently degrades performance compared to the base run in which no expansion is done (Voorhees and  Towell and Voorhees Disambiguating Highly Ambiguous Words Hou 1993). However, query expansion by lexically related words can significantly improve retrieval effectiveness: additional experiments in which hand-selected Word-Net synonym sets were used as seeds for expansion improved retrieval performance by over 30% (Voorhees 1994b). Because the process used in hand-picking the seed synonym sets encompassed more than simple sense resolution--other considerations such as specificity of the sense and perceived usefulness of the concept also played a part--simply finding the correct sense of the query terms is not likely to produce this large an improvement. Nonetheless, significant improvement should be possible if the correct sense can be determined.</Paragraph>
    <Paragraph position="8"> Unfortunately, determining the correct sense of a query word using simply the paradigmatic relations that organize WordNet and other thesauri is unlikely to be successful (Voorhees 1993). 1 Instead, the word sense disambiguation literature strongly suggests that syntagmatic relations are important for sense resolution. For example, consider the word board and the noun hierarchy of WordNet. Each of nail, hammer, and carpenter is a good clue for the 'lumber' sense of board, but each is closest to some other sense of board in WordNet when distance is measured by the number of IS-A links between the respective nodes.</Paragraph>
    <Paragraph position="9"> An ideal lexical system would therefore incorporate both paradigmatic and syntagmatic relations. An automatic text retrieval system could exploit such a combined lexical system by first using the syntagmatic relations to resolve word senses, and then adding both paradigmatic- and syntagmatic-related words to the query. The WordNet expansion experiments discussed above (Voorhees 1994b) suggest that paradigmaticrelated words are useful for expansion, while the success of retrieval techniques such as relevance feedback (Salton and Buckley 1990) demonstrates the usefulness of expansion by syntagmatic-related words. Since the different relations link quite different sets of words, the combined effect should be complimentary, resulting in greater improvement than either type of expansion alone.</Paragraph>
    <Paragraph position="10"> To test this conjecture, we must build a lexical system that contains both types of relations. This in turn requires capturing the syntagmatic relations associated with the various senses contained within a particular paradigmatic lexicon. A word sense disambiguator that can capture these relations is described in the remainder of the paper.</Paragraph>
  </Section>
class="xml-element"></Paper>