Towards a unified framework for bilingual terminology extraction of single-word and multi-word terms

Proceedings of the 27th International Conference on Computational Linguistics, pages 2855–2866 Santa Fe, New Mexico, USA, August 20-26, 2018. 2855 Towards a unified framework for bilingual terminology extraction of single-word and multi-word terms Jingshu Liu1,2, Emmanuel Morin1, and Sebasti´an Pe˜na Saldarriaga2 1LS2N - UMR CNRS 6004, Universit´e de Nantes, France 2Dictanova, Nantes, France 1{jingshu.liu, emmanuel.morin}@ls2n.fr 2{jingshu, spenasaldarriaga}@dictanova.com Abstract Extracting a bilingual terminology for multi-word terms from comparable corpora has not been widely researched. In this work we propose a unified framework for aligning bilingual terms independently of the term lengths. We also introduce some enhancements to the context-based and the neural network based approaches. Our experiments show the effectiveness of our en- hancements over previous works and that the system can be adapted in specialized domains. Title and Abstract in French Vers un syst`eme unifi´e pour l’extraction terminologique bilingue de termes simples et complexes L’extraction d’une terminologie bilingue pour les termes complexes `a partir de corpus compa- rables n’a pas ´et´e beaucoup ´etudi´ee. Dans ce travail, nous proposons un syst`eme unifi´e pour l’alignement des termes bilingues ind´ependamment de la longueur des termes. De plus nous in- troduisons ´egalement des am´eliorations aux approches bas´ees sur l’alignement de contexte et sur un r´eseau neuronal. Nos exp´eriences montrent l’efficacit´e de nos am´eliorations sur les travaux ant´erieurs, et le fait que le syst`eme peut ˆetre adapt´e en domaines de sp´ecialit´e. 1 Introduction Bilingual terminology extraction from comparable corpora has aroused a lot of attention since the 1990s (Fung, 1995; Rapp, 1999). Two classes of approach have been developed depending on the nature of the term to be aligned. The first class concerns the alignment of single-word terms using context-based or neural network approaches while the second attempts to align multi-word terms relying on compositional approaches. Few studies have focused on providing a unified framework for aligning single-word and multi-word terms, apart from Delpech et al. (2012) and Taslimipoor et al. (2016). The first requires some specific linguistic information like the morpheme translation table, which makes it difficult to employ for two languages from different linguistic families such as English and Chinese. The second incorporates word embeddings into their collocation alignment system but it is limited to several types of mapping that must match a set of pre-defined syntactic patterns. Our objective is to provide such a unified framework for aligning terms of variable length in specialized domains without specific linguistic knowledge of the source or target language. Large size comparable corpora in specialized domains are not always available. Consequently, many data driven systems cannot learn from enough information. A possible solution to this problem is to associate external data such as general domain corpora (Hazem and Morin, 2016) to the specialized corpora. Our work adapts this method in order to improve the system performance. Besides the enrichment of the data, our work focuses on studying and improving different state-of-the- art approaches both for single-word term and multi-word term alignments, which are finally unified into a single framework for alignment of terms of any length. …

Content truncated to 500 words