For our second participation to the Robust Word Sense Disambiguation (WSD) Task, we focused on performing a deep analysis of the ambiguity issue in the field of Information Retrieval. During the 2008 edition, we noted that although the WSD corpus allowed lifting lexical ambiguities, our results based on the corpus' WSD were not clearly better than those based on words only. We showed that lexical ambiguity was an issue only when queries included only one or possibly two words, but whenever the query was "longer", its words created a context that implicitly decreased lexical ambiguities. We thought we had a domain ambiguity problem, i.e. the retrieved documents did contain some of the query's words but they turned out to be irrelevant. Thus, we tried to expand the query's vocabulary in the following way: On the basis of the query's titles, we queried the Web (using Google's Search Engine) and selected the 50 top retrieved documents; We downloaded those documents and kept only the text; Then we trained a supervised classifier by associating the document classes to the query numbers; Finally we extracted the 50 most classifying words for each query.
2) 3) 4)
For an easier analysis of the answers, we converted the results into hypertext, thus allowing for a quick access to the text of the referenced documents. A detailed analysis of the answers showed that the ambiguity was in fact of semantic nature. The "right" or "wrong" documents were not differentiated by the words they contained: both included words from the query and from the domain. Thus the WHAT aspect (the topic) was equivalent. However, the HOW aspect (how people talked about the topic) was different and required a semantic “understanding” of the text. For instance, in the query dealing with the fourth victory of Indurain in the Tour de France (we were looking for documents relating to the reactions to this victory), all the answers were linked to the victory but some of them related to its anticipation while others were referring to it after it occurred. A human being can easily tell the "right" answers because of their experience of reactions to a victory in a bike contest. Therefore, in order to significantly improve the performance, we believe the problem should be addressed with methods allowing to introduce semantic elements.