Modelling Lexical Ambiguity in Computers, With and Without Word Senses

old_uid13740
titleModelling Lexical Ambiguity in Computers, With and Without Word Senses
start_date2014/04/01
schedule14h
onlineno
location_infosalle de conférences
summaryIn Computational Linguistics, there is a growing body of work that models word meaning from corpus data without recourse to a hand-crafted inventory. This has several advantages, including focusing computers on meanings which are attested in the data, but there is still the issue of whether word usages are best clustered into word senses or whether some other representation is better. In this talk I will describe a few projects using topic models and vector space models acquired from corpus data for representing word meaning. I will demonstrate how these models can be used both with and without an explicit representation of word sense. In applications such as novel sense detection for lexicography, representation of word sense is necessary, while for similarity judgements a more dynamic representation outperforms the alternatives. Linguists have advocated that a clear cut division into senses is not appropriate for all words, and that there is a spectrum of behaviour from clear cut cases of ambiguity to vague words with inter-related meanings. I will end with a brief introduction to current work which aims to identify empirically where on this spectrum a given lemma is.
responsiblesParoubek, Turner