Word Embedding Approach for Synonym Extraction of Multi-Word Terms

Word Embedding Approach for Synonym Extraction of Multi-Word Terms Amir Hazem and B´eatrice Daille Laboratoire des Sciences du Num´erique de Nantes (LS2N) Universit´e de Nantes, 44322 Nantes Cedex 3, France Amir.Hazem@univ-nantes.fr,Beatrice.Daille@univ-nantes.fr Abstract The acquisition of synonyms and quasi-synonyms of multi-word terms (MWTs) is a relatively new and under represented topic of research. However, dealing with MWT synonyms and semantically related terms is a challenging task, especially when MWT synonyms are single word terms (SWTs) or MWTs of different lengths. While several researches addressed synonym extraction of SWTs, few of them dealt with MWTs and fewer or none while MWTs synonyms are of variable lengths. The present research aims at introducing a new word-embedding-based approach for the automatic acquisition of synonyms of MWTs that manage length variability. We evaluate our approach on two specialized domain corpora, a French/English corpus of the wind energy domain and a French/English corpus of the breast cancer domain and show superior results compared to baseline approaches. Keywords: Synonym extraction, Multi-word terms, Compositionality, Word embeddings 1. Introduction Synonyms acquisition has mainly concerned single word terms (SWTs) using a variety of approaches such as: lexicon-based approaches (Blondel and Senellart, 2002), multilingual approaches (Wu and Zhou, 2003; van der Plas and Tiedemann, 2006; Andrade et al., 2013), distributional approaches (Lin, 1998; Hagiwara, 2008), etc. However, ex- ploring multi-word terms (MWTs) and their synonyms or semantically related terms can be useful in many applica- tions such as: word sense disambiguation, machine transla- tion, information retrieval, text simplification, etc. MWTs are motivated combinations that clearly convey the con- cept they designate. The requirement of term transparency argues in favor of compositional semantics for complex terms. Compositionality means that the whole meaning can be deduced from the meaning of its components and the syntactic rule by which they are combined (Partee et al., 1990). Pirrelli et al. (2010) claim that the most produc- tive compounds are compositional (at least weak composi- tional) constructions. Synonymic variants of multi-words exhibit multiple phenomena ranging from compositional multi-word terms synonyms of the same length such as: wind turbine/wind machine1; MWT synonyms of variable length such as: wind farm/wind power plant; to non com- positional MWT synonyms such as: pole tower/mast. Few works addressed the acquisition of MWT synonyms. The main approaches that have been proposed in the exper- imental literature deal with the acquisition of synonyms of MWTs that are compositional and often of the same length. Synonym extraction approaches implement the principle of compositionality by substituting parts of the MWT by syn- onyms provided by a dictionary (Hamon and Nazarenko, 2001), or by distributional analysis (Hazem and Daille, 2014). It has been recently shown that words, phrases, sentences, paragraphs and more generally, pieces of texts of any length can be efficiently represented by word embeddings using operations on vectors and matrices like addition or multipli- 1In the renewable energy domain. cation (Mitchell and Lapata, 2010; Mikolov et al., 2013b; Socher et al., 2011; Mikolov et al., 2013b; Le and Mikolov, 2014; Kalchbrenner et al., 2014; …

Content truncated to 500 words