Semantic Distances for Sets of Senses and Applications in Word Sense Disambiguation Dimitrios Mavroeidis, George Tsatsaronis and Michalis Vazirgiannis Department of Informatics, Athens University of Economics and Bussiness, Athens, Greece {dmavr,gbt,mvazirg}@aueb.gr Abstract. There has been an increasing interest both from the Information Retrieval community and the Data Mining community in investigating possible advantages of using Word Sense Disambiguation (WSD) for enhancing semantic information in the Information Retrieval and Data Mining process. Although contradictory results have been reported, there are strong indications that the use of WSD can contribute to the performance of IR and Data Mining algorithms. In this paper we propose two methods for calculating the semantic distance of a set of senses in a hierarchical thesaurus and utilize them for performing unsupervised WSD. Initial experiments have provided us with encouraging results. 1. Introduction Towards the direction of improving the accuracy in the retrieval process, the information retrieval community has been investigating the possible advantages of using Word Sense Disambiguation (WSD) [1] for enhancing both the query and the content with semantics. In spite of the early discouraging results [2], recent studies have clearly indicated that WSD algorithms achieving an accuracy of 50-60% can improve significantly the precision of IR tasks [3,4]. More precise experimental efforts [5] have even reported an absolute increase of 1.73% and a relative increase of 45.9% in precision whilst utilizing a supervised WSD algorithm that reported an accuracy of 62.1%. From the Data Mining Community perspective, the process of applying WSD for improving clustering or classification results has produced contradictory results. In [6,7] the results presented were negative, though probably because in [7] the WSD process applied did not assign a single sense to each word, but tackled all the possible senses for all the words, while in [6] the semantic relations, like the hypernym/hyponym relation, were not taken into account. In contrast, in [8,9], a rich representation for senses was utilized, that exploited the semantic relations between senses, as provided by WordNet [10]. Thus, there exist indications that the correct usage of senses can improve accuracy in Data Mining tasks. In general a WSD process can be either supervised or unsupervised (or a combination of the two). The supervised WSD considers a pre-tagged text corpus that is used as a training set. The sense of a new keyword can then be inferred based on