Integrating Distributional and Lexical Information for Semantic Classification of Words using MRMF

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 2708–2717, Osaka, Japan, December 11-17 2016. Integrating Distributional and Lexical Information for Semantic Classification of Words using MRMF Rosa Tsegaye Aga Hochschule Hannover Hanover, Germany rosa-tsegaye.aga@hs-hannover.de Lucas Drumond Universität Hildesheim Hildesheim, Germany ldrumond@gmail.com Christian Wartena Hochschule Hannover Hanover, Germany christian.wartena@hs-hannover.de Lars Schmidt-Thieme Universität Hildesheim Hildesheim, Germany schmidt-thieme@ismll.de Abstract Semantic classification of words using distributional features is usually based on the semantic similarity of words. We show on two different datasets that a trained classifier using the distribu- tional features directly gives better results. We use Support Vector Machines (SVM) and Multi- relational Matrix Factorization (MRMF) to train classifiers. Both give similar results. However, MRMF, that was not used for semantic classification with distributional features before, can eas- ily be extended with more matrices containing more information from different sources on the same problem. We demonstrate the effectiveness of the novel approach by including informa- tion from WordNet. Thus we show, that MRMF provides an interesting approach for building semantic classifiers that (1) gives better results than unsupervised approaches based on vector similarity, (2) gives similar results as other supervised methods and (3) can naturally be extended with other sources of information in order to improve the results. 1 Introduction In this paper we consider the task of classifying words into a large number of semantic categories. For this, we use two different data sets: 1. A dataset which is used in literature (Bullinaria and Levy, 2007) - to enable compare our results with results reported in the literature, 2. A larger dataset that is derived from a large thesaurus. The second dataset comes close to practical applications for semantic word classification. Organizations maintaining thesauri usually try to keep their thesaurus up to date and frequently add new terminology to the thesaurus. For each new term, they have to decide at what point it has to be inserted. Automatic semantic classification supports exactly this task. However, the classifier should be able to choose from hundreds or even thousands of semantic classes, not just from a dozen. For semantic word classification, it is a common approach to represent words by context features. Usually, co-occurrence statistics are used as context features. According to the distributional hypothesis, words with similar context features have a similar meaning. Thus, we can use any distance measure between the feature vectors as a measure of semantic similarity. These distances are now commonly used in a nearest neighbor or a nearest centroid (or nearest prototype) classifier. Recently, distributional features have also been used directly to train classifiers that classify pairs of words as being synonymous or not (Hagiwara, 2008; Weeds et al., 2014; Aga et al., 2016). In these approaches, first a vector repre- sentation for a pair is build, that is used by a machine learning algorithm. In the following we will use the distributional features directly to categorize the words into a large number of categories. We will also see that the supervised …

Content truncated to 500 words