Gloss-Based Semantic Similarity Metrics for Predominant Sense Acquisition

Gloss-Based Semantic Similarity Metrics for Predominant Sense Acquisition Ryu Iida Nara Institute of Science and Technology 8916-5 Takayama, Ikoma, Nara, 630-0192, Japan ryu-i@is.naist.jp Diana McCarthy and Rob Koeling University of Sussex Falmer, East Sussex BN1 9QH, UK {dianam,robk}@sussex.ac.uk Abstract In recent years there have been various ap- proaches aimed at automatic acquisition of predominant senses of words. This infor- mation can be exploited as a powerful back- off strategy for word sense disambiguation given the zipfian distribution of word senses. Approaches which do not require manually sense-tagged data have been proposed for English exploiting lexical resources avail- able, notably WordNet. In these approaches distributional similarity is coupled with a se- mantic similarity measure which ties the dis- tributionally related words to the sense in- ventory. The semantic similarity measures that have been used have all taken advantage of the hierarchical information in WordNet. We investigate the applicability to Japanese and demonstrate the feasibility of a mea- sure which uses only information in the dic- tionary definitions, in contrast with previ- ous work on English which uses hierarchi- cal information in addition to dictionary def- initions. We extend the definition based semantic similarity measure with distribu- tional similarity applied to the words in dif- ferent definitions. This increases the recall of our method and in some cases, precision as well. 1 Introduction Word sense disambiguation (WSD) has been an ac- tive area of research over the last decade because many researches believe it will be important for applications which require, or would benefit from, some degree of semantic interpretation. There has been considerable skepticism over whether WSD will actually improve performance of applications, but we are now starting to see improvement in per- formance due to WSD in cross-lingual information retrieval (Clough and Stevenson, 2004; Vossen et al., 2006) and machine translation (Carpuat and Wu, 2007; Chan et al., 2007) and we hope that other ap- plications such as question-answering, text simplifi- cation and summarisation might also benefit as WSD methods improve. In addition to contextual evidence, most WSD sys- tems exploit information on the most likely mean- ing of a word regardless of context. This is a pow- erful back-off strategy given the skewed nature of word sense distributions. For example, in the En- glish coarse grained all words task (Navigli et al., 2007) at the recent SemEval Workshop the base- line of choosing the most frequent sense using the first WordNet sense attained precision and recall of 78.9% which is only a few percent lower than the top scoring system which obtained 82.5%. This finding is in line with previous results (Snyder and Palmer, 2004). Systems using a first sense heuristic have relied on sense-tagged data or lexicographer judg- ment as to which is the predominant sense of a word. However sense-tagged data is expensive and further- more the predominant sense of a word will vary de- pending on the domain (Koeling et al., 2005; Chan and Ng, 2007). One direction of research following McCarthy et al. (2004) has …

Content truncated to 500 words