Detecting Polysemy in Hard and Soft Cluster Analyses of German Preposition Vector Spaces

International Joint Conference on Natural Language Processing, pages 632–640, Nagoya, Japan, 14-18 October 2013. Detecting Polysemy in Hard and Soft Cluster Analyses of German Preposition Vector Spaces Sylvia Springorum, Sabine Schulte im Walde and Jason Utt Institut f¨ur Maschinelle Sprachverarbeitung Universit¨at Stuttgart Pfaffenwaldring 5b, 70569 Stuttgart, Germany {riestesa,schulte,uttjn}@ims.uni-stuttgart.de Abstract This paper presents a methodology to identify polysemous German prepositions by exploring their vector spatial proper- ties. We apply two cluster evaluation metrics (the Silhouette Value (Kaufman and Rousseeuw, 1990) and a fuzzy ver- sion of the V-Measure (Rosenberg and Hirschberg, 2007)) as well as various cor- relations, to exploit hard vs. soft cluster analyses based on Self-Organising Maps. Our main hypothesis is that polysemous prepositions are outliers, and thus repre- sent either (i) singletons or (ii) marginals of the clusters within a cluster analysis. Our analyses demonstrate that (a) in a sub- set of the clusterings, singletons have a tendency to contain polysemous preposi- tions; and (b) misclassification and cluster membership rate exhibit a moderate corre- lation with ambiguity rate. 1 Introduction Vector space models have become a steadily in- creasing, integral part of data-intensive lexical se- mantics over the past 20 years (cf. Turney and Pantel (2010) and Erk (2012) for two recent sur- veys). They have been exploited in psycholinguis- tic (Lund and Burgess, 1996) and computational linguistic research (Sch¨utze, 1998), to explore dis- tributional properties of target objects and the no- tion of “similarity” within a geometric setting. While individual vector space approaches have been concerned with sense discrimination, it is still largely unknown how to identify polyse- mous objects within a vector space model, and which geometric properties characterise the poly- semous objects. For example, Sch¨utze (1998) per- formed sense discrimination of ambiguous word tokens, based on their second-order co-occurrence distributions; Erk (2009) presented two variants of defining regions of word meaning in vector spaces; Erk and Pad´o (2010) defined a model where polysemous words activated several word vectors; Boleda et al. (2012b) compared two mod- els of representing regular polysemy, one with multiple class assignments for multiple senses, and one incorporating classes with polysemy properties. Our work is different from all these ap- proaches, since we aim to investigate prototypical spatial properties of polysemous objects. More specifically, this paper is part of a larger framework that systematically explores the vec- tor spatial properties of German prepositions, a notoriously polysemous closed word class. Re- lying on Self-Organising Maps (SOMs, cf. Ko- honen (2001)) and preposition-dependent nouns as vector-space features, we present a method- ology to identify the degree of polysemy of the prepositions. For this task, the methodology ap- plies two cluster evaluation metrics, the Silhou- ette Value (Kaufman and Rousseeuw, 1990) and the V-Measure (Rosenberg and Hirschberg, 2007), to hard vs. soft cluster analyses based on the Self- Organising Maps. Since we start out with a hard clustering, a sub-task is concerned with transfer- ring the SOM hard clusters to soft clusters. Simi- larly, the original V-Measure applies to hard clus- ters only, so a …

Content truncated to 500 words