Visualisation and Exploration of High-Dimensional Distributional Features in Lexical Semantic Classification

Visualisation and Exploration of High-Dimensional Distributional Features in Lexical Semantic Classification Maximilian K¨oper1, Melanie Zaiß2, Qi Han2, Steffen Koch2, Sabine Schulte im Walde1 1Institut f¨ur Maschinelle Sprachverarbeitung 2Institut f¨ur Visualisierung und Interaktive Systeme Universit¨at Stuttgart, Germany 1 {maximilian.koeper,schulte}@ims.uni-stuttgart.de 2 {firstname.lastname}@vis.uni-stuttgart.de Abstract Vector space models and distributional information are widely used in NLP. The models typically rely on complex, high-dimensional objects. We present an interactive visualisation tool to explore salient lexical-semantic features of high-dimensional word objects and word similarities. Most visualisation tools provide only one low-dimensional map of the underlying data, so they are not capable of retaining the local and the global structure. We overcome this limitation by providing an additional trust-view to obtain a more realistic picture of the actual object distances. Additional tool options include the reference to a gold standard classification, the reference to a cluster analysis as well as listing the most salient (common) features for a selected subset of the words. Keywords: vectors space models, high-dimensional features, interactive visualisation 1. Introduction Vector space models (VSMs) and distributional information have been a steadily increasing, integral part of lexical se- mantic research over the past 20 years (Turney and Pantel, 2010): Relying on the distributional hypothesis that words which occur in similar contexts tend to have similar mean- ings (Harris, 1954; Firth, 1957), VSMs have been exploited to determine co-occurrence features of target words, and to explore the notion of “similarity” between target words within a geometric setting. While the emergence of high-dimensional vector space mod- els is increasingly pervasive within data-intensive lexical semantics, and even though useful features have been identi- fied in general terms:1 when it comes to a specific semantic phenomenon, we need to explore the semantically salient distributional features in order to investigate the respective phenomenon. Our research focuses on lexical semantic classification: au- tomatic classification of words according to their meaning. The lexical classifications are typically related to the word class, i.e., we are interested in identifying semantic classes of verbs, or nouns, or prepositions, etc. For example, compu- tational semantic classes of nouns might resemble WordNet synsets (Miller et al., 1990); semantic classes of verbs might resemble FrameNet or VerbNet classes (Fillmore et al., 2003; Kipper Schuler, 2006); semantic classes of prepositions might resemble the Preposition Project classes (Litkowski and Hargraves, 2005). We aim for a theoretically and cogni- tively adequate selection of distributional features to model word meaning and word relatedness, i.e., our goal is to op- timise the choice of high-dimensional salient features that best describe and distinguish the word meanings and the semantic classes of the words. This paper presents an interactive visual exploration tool that supports our high-dimensional models of lexical- 1See Agirre et al. (2009) and Bullinaria and Levy (2007; 2012), among others, for systematic comparisons of co-occurrence fea- tures on various semantic relatedness tasks. semantic objects, features and classes. Based on high- dimensional word co-occurrence features, the tool provides low-dimensional maps of the underlying objects enhanced by a trust-view to obtain a more realistic picture …

Content truncated to 500 words