Exploring Vector Space Models to Predict the Compositionality of German Noun-Noun Compounds

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 1: Proceedings of the Main Conference and the Shared Task, pages 255–265, Atlanta, Georgia, June 13-14, 2013. c⃝2013 Association for Computational Linguistics Exploring Vector Space Models to Predict the Compositionality of German Noun-Noun Compounds Sabine Schulte im Walde and Stefan M¨uller and Stephen Roller Institut f¨ur Maschinelle Sprachverarbeitung Universit¨at Stuttgart Pfaffenwaldring 5b, 70569 Stuttgart, Germany {schulte,muellesn,roller}@ims.uni-stuttgart.de Abstract This paper explores two hypotheses regarding vector space models that predict the compo- sitionality of German noun-noun compounds: (1) Against our intuition, we demonstrate that window-based rather than syntax-based distri- butional features perform better predictions, and that not adjectives or verbs but nouns rep- resent the most salient part-of-speech. Our overall best result is state-of-the-art, reach- ing Spearman’s ρ = 0.65 with a word- space model of nominal features from a 20- word window of a 1.5 billion word web cor- pus. (2) While there are no significant dif- ferences in predicting compound–modifier vs. compound–head ratings on compositionality, we show that the modifier (rather than the head) properties predominantly influence the degree of compositionality of the compound. 1 Introduction Vector space models and distributional information have been a steadily increasing, integral part of lex- ical semantic research over the past 20 years. On the one hand, vector space models (see Turney and Pantel (2010) and Erk (2012) for two recent sur- veys) have been exploited in psycholinguistic (Lund and Burgess, 1996) and computational linguistic re- search (Sch¨utze, 1992) to explore the notion of “sim- ilarity” between a set of target objects within a ge- ometric setting. On the other hand, the distribu- tional hypothesis (Firth, 1957; Harris, 1968) has been exploited to determine co-occurrence features for vector space models that best describe the words, phrases, sentences, etc. of interest. While the emergence of vector space models is in- creasingly pervasive within data-intensive lexical se- mantics, and even though useful features have been identified in general terms:1 when it comes to a spe- cific semantic phenomenon, we need to explore the relevant distributional features in order to investigate the respective phenomenon. Our research is inter- ested in the meaning of German compounds. More specifically, we aim to predict the degrees of compo- sitionality of German noun-noun compounds (e.g., Feuerwerk ‘fire works’) with regard to the mean- ings of their constituents (e.g., Feuer ‘fire’ and Werk ‘opus’). This prediction uses vector space models, and our goal is to identify salient features that de- termine the degree of compositionality of the com- pounds by relying on the distributional similarities between the compounds and their constituents. In this vein, we systematically explore window- based and syntax-based contextual clues. Since the targets in our vector space models are all nouns (i.e., the compound nouns, the modifier nouns, and the head nouns), our hypothesis is that adjectives and verbs are expected to provide salient distri- butional properties, as adjective/verb meaning and noun meaning are in a strong interdependent rela- tionship. Even more, we expect adjectives and verbs that are …

Content truncated to 500 words