Improving distributional thesauri by exploring the graph of neighbors

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 709–720, Dublin, Ireland, August 23-29 2014. Improving distributional thesauri by exploring the graph of neighbors Vincent Claveau IRISA - CNRS Campus de Beaulieu 35042 Rennes, France vincent.claveau@irisa.fr Ewa Kijak IRISA - Univ. of Rennes 1 Campus de Beaulieu 35042 Rennes, France ewa.kijak@irisa.fr Olivier Ferret CEA, LIST LVIC 91191 Gif-sur-Yvette, France olivier.ferret@cea.fr Abstract In this paper, we address the issue of building and improving a distributional thesaurus. We first show that existing tools from the information retrieval domain can be directly used in order to build a thesaurus with state-of-the-art performance. Secondly, we focus more specifically on improving the obtained thesaurus, seen as a graph of k-nearest neighbors. By exploiting information about the neighborhood contained in this graph, we propose several contributions. 1) We show how the lists of neighbors can be globally improved by examining the reciprocity of the neighboring relation, that is, the fact that a word can be close of another and vice-versa. 2) We also propose a method to associate a confidence score to any lists of nearest neighbors (i.e. any entry of the thesaurus). 3) Last, we demonstrate how these confidence scores can be used to reorder the closest neighbors of a word. These different contributions are validated through experiments and offer significant improvement over the state-of-the-art. 1 Introduction Distributional thesauri are useful for many NLP tasks and their construction is an issue widely discussed for several years (Grefenstette, 1994). However this is still a very active research field, maintained by the increasingly large number of available corpus and by many applications. These thesauri associate each of their entry with a list of words that are desired semantically close to the entry. This notion of proximity varies (synonymy, other paradigmatic relations, syntagmatic relations (Budanitsky and Hirst, 2006; Adam et al., 2013, for a discussion)), but the methods used for the automatic construction of thesauri are often shared. For the most part, these methods rely on the distributional hypothesis of (Firth, 1957): each word is characterized by the set of contexts in which it appears, and the semantic proximity of two words can be inferred from the proximity of their contexts. This hypothesis has been implemented in different ways, and several propositions to improve the results have been explored (see next section for a state of the art). The work presented in this article are part of this framework. We propose several contributions on the creation of these distributional thesauri and their improvement. We first show that models from information retrieval (IR) can provide information on semantic relationships, and are thus adapted to the task of creating these thesauri. In addition, they offer very competitive results compared to the state of the art, while enjoying existing tools (Section 3). The most important part of our work then focuses on the exploitation of such semantic neighborhood relations. The IR models indeed provide lists ordering all words by decreasing similarity, that form a graph of …

Content truncated to 500 words