HsH: Estimating Semantic Similarity of Words and Short Phrases with Frequency Normalized Distance Measures

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 2: Seventh International Workshop on Semantic Evaluation (SemEval 2013), pages 48–52, Atlanta, Georgia, June 14-15, 2013. c⃝2013 Association for Computational Linguistics HsH: Estimating Semantic Similarity of Words and Short Phrases with Frequency Normalized Distance Measures Christian Wartena Hochschule Hannover – University of Applied Sciences and Arts Department of Information and Communication Expo Plaza 12, 30539 Hannover, Germany Christian.Wartena@hs-hannover.de Abstract This paper describes the approach of the Hochschule Hannover to the SemEval 2013 Task Evaluating Phrasal Semantics. In or- der to compare a single word with a two word phrase we compute various distributional sim- ilarities, among which a new similarity mea- sure, based on Jensen-Shannon Divergence with a correction for frequency effects. The classification is done by a support vector ma- chine that uses all similarities as features. The approach turned out to be the most successful one in the task. 1 Introduction The task Evaluating Phrasal Semantics of the 2013 International Workshop on Semantic Evaluation (Manandhar and Yuret, 2013) consists of two sub- tasks. For the first subtask a list of pairs consisting of a single word and a two word phrase are given. For the English task a labeled list of 11,722 pairs was provided for training and a test set with 3,906 unlabeled examples. For German the training set contains 2,202 and the test set 732 pairs. The system should be able to tell whether the two word phrase is a definition of the single word or not. This task is somewhat different from the usual perspective of finding synonyms, since definitions are usually more general than the words they define. In distributional semantics words are represented by context vectors and similarities of these con- text vectors are assumed to reflect similarities of the words they represent. We compute context vec- tors for all words using the lemmatized version of the Wacky Corpora for English (UKWaC, approxi- mately 2,2 billion words) and German (DeWaC, 1,7 billion words) (Baroni et al., 2009). For the phrases we compute the context vectors as well directly on the base of occurrences of that phrase, as well as by construction from the context vectors of the two components. For the similarities between the vec- tors we use Jensen-Shannon divergence (JSD) and cosine similarity. Since the JSD is extremely depen- dent on the number of occurrences of the words, we define a new similarity measure that corrects for this dependency. Since none of the measures gives satis- factory results, we use all measures to train a support vector machine that classifies the pairs. The remainder of this paper is organized as fol- lows. We start with an overview of related work. In section 3 we discuss the dependence of JSD on word frequency and introduce a new similarity measure. Section 4 then describes the system. The results are given in section 5 and are discussed in section 6. 2 Related Work Though distributional similarity has widely been studied and has become an established …

Content truncated to 500 words