Looking at word meaning. An interactive visualization of Semantic Vector Spaces for Dutch synsets

Proceedings of the EACL 2012 Joint Workshop of LINGVIS & UNCLH, pages 16–24, Avignon, France, April 23 - 24 2012. c⃝2012 Association for Computational Linguistics Looking at word meaning. An interactive visualization of Semantic Vector Spaces for Dutch synsets Kris Heylen, Dirk Speelman and Dirk Geeraerts QLVL, University of Leuven Blijde-Inkomsstraat 21/3308, 3000 Leuven (Belgium) {kris.heylen, dirk.speelman, dirk.geeraerts}@arts.kuleuven.be Abstract In statistical NLP, Semantic Vector Spaces (SVS) are the standard technique for the automatic modeling of lexical semantics. However, it is largely unclear how these black-box techniques exactly capture word meaning. To explore the way an SVS struc- tures the individual occurrences of words, we use a non-parametric MDS solution of a token-by-token similarity matrix. The MDS solution is visualized in an interac- tive plot with the Google Chart Tools. As a case study, we look at the occurrences of 476 Dutch nouns grouped in 214 synsets. 1 Introduction In the last twenty years, distributional models of semantics have become the standard way of mod- eling lexical semantics in statistical NLP. These models, aka Semantic Vector Spaces (SVSs) or Word Spaces, capture word meaning in terms of frequency distributions of words over co- occurring context words in a large corpus. The basic assumption of the approach is that words occurring in similar contexts will have a simi- lar meaning. Speficic implementations of this general idea have been developed for a wide va- riety of computational linguistic tasks, includ- ing Thesaurus extraction and Word Sense Dis- ambiguation, Question answering and the model- ing of human behavior in psycholinguistic experi- ments (see Turney and Pantel (2010) for a general overview of applications and speficic models). In recent years, Semantic Vector Spaces have also seen applications in more traditional domains of linguistics, like diachronic lexical studies (Sagi et al., 2009; Cook and Stevenson, 2010; Rohrdantz et al., 2011) , or the study of lexical variation (Peirsman et al., 2010). In this paper, we want to show how Semantic Vector Spaces can further aid the linguistic analysis of lexical semantics, pro- vided that they are made accessible to lexicolo- gists and lexicographers through a visualization of their output. Although all applications mentioned above as- sume that distributional models can capture word meaning to some extent, most of them use SVSs only in an indirect, black-box way, without an- alyzing which semantic properties and relations actually manifest themselves in the models. This is mainly a consequence of the task-based evalu- ation paradigm prevalent in Computational Lin- guistics: the researchers address a specific task for which there is a pre-defined gold standard; they implement a model with some new features, that usually stem from a fairly intuitive, common- sense reasoning of why some feature might bene- fit the task at hand; the new model is then tested against the gold standard data and there is an eval- uation in terms of precision, recall and F-score. In rare cases, there is also an error analysis that leads to hypotheses about semantic characteristics that are not yet properly modeled. …

Content truncated to 500 words