The Senseval-2 exercise provided an unprece­dented opportunity to explore word sense dis­ambiguation (WSD) in a common evaluation framework for a large number of languages. In past work, we have focused on unsupervised methods for English, taking advantage of the WordNet hierarchy and sometimes also selec-tional preferences between predicates and argu­ments (Resnik, 1997; Resnik, 1999). In the cur­rent exercise, however, WordNet-like sense hi­erarchies were not necessarily going to be avail­able for all languages, and the predominance of lexical selection tasks (rather than all-words tasks) suggested adopting a disambiguation ap­proach capable of exploiting manually anno­tated training data. These considerations mo­tivated a system design based on supervised learning, where senses to be predicted did not need to be treated as part of a semantic hierar­chy.

Our design was also motivated by the role of semantic selection techniques in our longer term research agenda. In the context of our group's work on cross-language information retrieval and machine translation applications (Resnik et al., 2001; Cabezas et al., 2001), lexical selection — that is, choosing the right target-language word given a source-language word in context — is a crucial task. Because the lexical selec­tion problem is extremely similar to sense selec­tion, and because this was our first foray into supervised methods, we took advantage of the opportunity to construct an architecture that will support both tasks.

In the sections that follow, we lay out our system architecture, briefly summarize our Senseval-2 results, and discuss our plans for future work.

UMD's system follows the classic supervised learning paradigm that, for WSD, is perhaps best exemplified by Yarowsky's (1993) work. Each word in the vocabulary is considered an independent classification problem. First, an­notated training instances for the ambiguous word are analyzed so that each instance can be represented as a collection of feature-value pairs labeled with the correct category. Then, these data are used for parameter estimation within a supervised learning framework in or­der to produce a trained classifier. Finally, the trained classifier is given previously unseen test instances and for each instance it predicts what the appropriate category label should be.

Table 1 shows the performance of UMD's su­pervised sense tagger (UMD-SST) for the lex­ical sample tasks in four languages. The fig­ures for English, Spanish, and Swedish are offi­cial Senseval-2 results; the figures for Basque are unofficial results kindly computed by the Basque task organizers after Senseval-2 be­cause our Basque responses were not submitted in time for official evaluation.

In general, we were quite pleased with the re­sults, particularly since this was our first time participating in Senseval. UMD-SST turned in a solid performance in comparison with the baselines and other systems, with essentially no language-specific alterations necessary other than those required for tokenization. This en­abled us to participate in system evaluation for more languages than any site except JHU. We consider this a good starting point for our fur­ther investigations, which we now briefly de­scribe.

Using the current system as a starting point, we are engaged in three lines of further investi­gation: linguistically richer contextual features, corpus-dependent expansion of feature vectors, and lexical selection via supervised learning.

In our preliminary tests using training and development data, we experimented first with using Pyjife as the feature set, and obtained sig­nificant improvements when we added ^co]iocin order to capture collocations and other local contextual features. In our follow-up efforts we plan to use broad-coverage parsing to create a set of features augmented further by grammat­ical relations, thus capturing collocations medi­ated by syntactic structure. For example, al­though our current feature vectors could not represent the presence of the word tagger as a nearby collocate of the word describe in the ab­stract of this paper, syntactically richer repre­sentations of this context for the verb describe would include the feature ob j ect=> t agger '. Use of syntactic collocates will require broad-coverage parsing in all the languages of inter­est in order to identify grammatical relations; for this we will take advantage of our other work at Maryland on bootstrapping stochastic parsers for new languages using parallel corpora (Cabezas et al., 2001).

In our preliminary efforts we were not sur­prised to find that sparseness of data was a problem. Although we expect that some improvements may be obtained by collapsing across word variants — e.g. via morphologi­cal equivalence classes or stemming — we also plan to focus our efforts on semantic expansion, using document expansion techniques we have developed in our research on cross-language in­formation retrieval (Levow et al., 2001). We have implemented a variant of the architecture in which training contexts are used as queries to a comparable corpus in order to retrieve re­lated documents. The features from these docu­ments are then added to the context representa­tions, providing semantically enhanced feature vectors. Evaluation of this approach using Sen­seval data is in progress.

Our third avenue of investigation focuses on the use of our supervised WSD infrastructure to address problems of lexical selection in ma­chine translation. Empirically, there is a close relationship between sense distinctions and pat­terns of lexicalization across languages (Resnik and Yarowsky, 1999). And operationally, there is no real difference between labeling a word with a sense tag from a monolingual dictionary and labeling that word with a translation from a bilingual dictionary. Using WSD techniques for lexical selection primarily requires solving two problems. The first problem is acquisition of annotated training data, and in this case large corpora of translation-labeled words in context can be created by obtaining parallel corpora, performing word-level alignment, and labeling each word with its correspondent in the other language; this problem is already solved as part of our infrastructure for research on statistical machine translation (Cabezas et al., 2001). The second problem is one of scalability: the ap­proach we have described requires a separate classifer for every sense (or, now, every possi­ble word-level translation) of every source lan­guage word. This remains an open issue, but we are optimistic about rapid developments in this area since scaling up to large vocabularies is a problem shared by everybody who wishes to use supervised WSD techniques in a broad-coverage setting.

University of Maryland's sense tagger repre­sents a classic instance of the supervised learn­ing approach. At the same time, we have made architectural choices that promote language in­dependence, modularity, extensibility, and scal­ability, and in a relatively short time period we succeeded in putting together an implementa­tion that performs quite credibly among an im­pressive collection of competitors. We are en­couraged by the results and we look forward to participating in further Senseval exercises.

This work was supported in part by De­partment of Defense contract MDA90496C1250 and DARPA/ITO Cooperative Agreement N660010028910. We're very grateful to all the Senseval-2 organizers and task organizers for their hard work, to Thorsten Joachims for mak­ing SVM-Light available, and to David Mar­tinez for computing our results for Basque.