Word Sense Disambiguation: Why Statistics When We Have These Numbers?

Word sense disambiguation continues to be a di cult problem in machine translation (MT). Current methods either demand large amounts of corpus data and training or rely on knowledge of hard selectional constraints. In either case, the methods have been demonstrated only on a small scale and mostly in isolation, where disambiguation is a task by itself. It is not clear that the methods can be scaled up and integrated with other components of analysis and generation that constitute an end-to-end MT system. In this paper, we illustrate how the Mikrokosmos Knowledge-Based MT system disambiguates word senses in real-world texts with a very high degree of correctness. Disambiguation in Mikrokosmos is achieved by a combination of (i) a broad-coverage ontology with many selectional constraints per concept, (ii) a large computational-semantic lexicon grounded in the ontology, (iii) an optimized search algorithm for checking selectional constraints in the ontology, and (iv) an e cient control mechanism with near-linear processing complexity. Moreover, Mikrokosmos constructs complete meaning representations of an input text using the chosen word senses. 1 Word Sense Ambiguity Word sense disambiguation continues to be a di cult problem for machine translation (MT) systems.The most common current methods for resolving word sense ambiguities are based on statistical collocations or static selectional preferences between pairs of word senses. The real power of word sense selection seems to lie in the ability to constrain the possible senses of a word based on selections made for other words in the local context. Although methods using selectional constraints and semantic networks have been delineated at least since Katz and Fodor (1963), computational models have not demonstrated the e ectiveness of knowledgebased methods in resolving word senses in real-world texts on a large scale. This has resulted in a predominant shift of attention from knowledge-based to corpus-based, statistical methods for word sense resolution, despite the far greater potential of knowledge-based methods for advancing the development of large, practical, domain independent NLP/MT systems. In this article, we illustrate how the semantic analyzer of the Mikrokosmos machine translation system resolves word sense ambiguities in real-world Spanish texts (news articles on company mergers and acquisitions from the EFE newswire) with a high degree of correctness. We begin by presenting the results from Mikrokosmos and then illustrate how they were obtained. 1 See Guthrie et al (1996) and Wilks et al (1995) for recent surveys of related work. Text #1 #2 #3 #4 Average # words 347 385 370 353 364 # words/sentence 16.5 24.0 26.4 20.8 21.4 # open-class words 183 167 177 177 176 # ambiguous open-class words 57 42 57 35 48 # resolved by syntax 21 19 20 12 18 total # correctly resolved 51 41 45 34 43 % correct 97% 99% 93% 99% 97% Table 1. Mikrokosmos Results in Disambiguating Open Class Words in Spanish Texts.