Learning Thesaurus Relations from Distributional Features

Learning Thesaurus Relations from Distributional Features Rosa Tsegaye Aga, Christian Wartena, Lucas Drumond, Lars Schmidt-Thieme Hochschule Hannover, Universitat Hildesheim rosa-tsegaye.aga@hs-hannover.de, christian.wartena@hs-hannover.de, ldrumond@ismll.de,schmidt-thieme@ismll.info Abstract In distributional semantics words are represented by aggregated context features. The similarity of words can be computed by comparing their feature vectors. Thus, we can predict whether two words are synonymous or similar with respect to some other semantic relation. We will show on six different datasets of pairs of similar and non-similar words that a supervised learning algorithm on feature vectors representing pairs of words outperforms cosine similarity between vectors representing single words. We compared different methods to construct a feature vector representing a pair of words. We show that simple methods like pairwise addition or multiplication give better results than a recently proposed method that combines different types of features. The semantic relation we consider is relatedness of terms in thesauri for intellectual document classification. Thus our findings can directly be applied for the maintenance and extension of such thesauri. To the best of our knowledge this relation was not considered before in the field of distributional semantics. Keywords: Distributional semantics, thesauri, context vectors, Supervised machine learning 1. Introduction In the past decades, distributional similarity (DS) has been used to solve many different tasks related to the meaning of words. The main idea of DS is, that words that appear in similar contexts are likely to have a similar meaning. When context representations are built from large amounts of texts, DS can give good results for various tasks. How- ever, DS does not directly correspond to any traditional se- mantic relation. Distributional similar words might be syn- onyms, antonyms, hypernyms etc. In this paper we will consider a very basic and central task for computational lexical semantics: the decision whether two words are semantically related or not. A simple ap- proach is to compute a similarity between the words and learn a threshold above which the words can be considered as being related. We show that a much stronger supervised approach, in which the similarity between the words itself is learned from examples, gives much better results. To do so we need to construct feature vectors representing pairs of words. Especially, we show that a simple combination of context vectors gives better results than using vectors com- bining different types of features. Much research on DS has been done on very small datasets; For example, the TOEFL data, presumably the most of- ten used dataset in this field, just has 80 questions. Be- cause we need large datasets, we use data sampled from large thesauri, especially Eurovoc (Office for Official Pub- lications of the European Communities, 1995) and STW (ZBW - Leibniz Information Centre for Economics, 2014). These data have two advantages: It is easy to sample large amounts of related (and non-related) words and the data are constructed independent of any specific semantic task. Since thesauri are also used for automatic indexing and for full-text retrieval, it is important to know all possible terms …

Content truncated to 500 words