Building Comparable Corpora for Assessing Multi-Word Term Alignment

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 3103–3112 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 3103 Building Comparable Corpora for Assessing Multi-Word Term Alignment Omar Adjali1, Emmanuel Morin2, Pierre Zweigenbaum1 1Universit´e Paris-Saclay, CNRS, Laboratoire Interdisciplinaire des Sciences du Num´erique, Orsay, France 2Nantes Universit´e, CNRS, Laboratoire des Sciences du Num´erique de Nantes, France omar.adjali@universite-paris-saclay.fr, pz@lisn.fr, emmanuel.morin@univ-nantes.fr Abstract Recent work has demonstrated the importance of dealing with Multi-Word Terms (MWTs) in several Natural Language Pro- cessing applications. In particular, MWTs pose serious challenges for alignment and machine translation systems because of their syntactic and semantic properties. Thus, developing algorithms that handle MWTs is becoming essential for many NLP tasks. However, the availability of bilingual and more generally multi-lingual resources is limited, especially for low-resourced languages and in specialized domains. In this paper, we propose an approach for building comparable corpora and bilingual term dictionaries that help evaluate bilingual term alignment in comparable corpora. To that aim, we exploit parallel corpora to perform automatic bilingual MWT extraction and comparable corpus construction. Parallel information helps to align bilingual MWTs and makes it easier to build comparable specialized sub-corpora. Experimental validation on an existing dataset and on manually annotated data shows the interest of the proposed methodology. 1. Introduction The compilation of bilingual terminological resources has become critical for many NLP tasks such as cross- lingual information retrieval (Miangah, 2008), machine translation (Arcan et al., 2014; Yang et al., 2016) and many others: bilingual terminologies help such tasks either by reducing their computational cost or by im- proving their performance. Acquiring bilingual ter- minological resources is a difficult task, as it requires tremendous manual annotation effort. Early research has been conducted on automatic approaches for the acquisition of bilingual terminologies (Kupiec, 1993; Daille et al., 1994; Vintar, 2001; Wu and Chang, 2004). Bilingual term extraction (BTE) approaches first iden- tify monolingual terms, and then establish cross-lingual correspondences between pairs of terms using align- ment methods. Two main approaches have emerged from the liter- ature, one that relies on comparable corpora (Rapp, 1995; Tanaka and Iwasaki, 1996; Fung, 1998; Chiao and Zweigenbaum, 2002; Otero, 2007; Morin et al., 2007; Saralegi et al., 2008; Fiˇser et al., 2011; Fiˇser and Ljubeˇsi´c, 2011; Ljubeˇsic et al., 2012; Aker et al., 2013; Hazem and Morin, 2016) while the other leverages information from parallel corpora (Somers, 2001; Kwong et al., 2004; Fan et al., 2009; Lefever et al., 2009; Macken et al., 2013; Arcan et al., 2014; Yang et al., 2016; Krstev et al., 2018; ˇSandrih et al., 2020). Comparable corpora include non-parallel texts in different languages that share similar characteristics. While compiling them is quite easy, for example from the web, comparable corpora-based BTE needs addi- tional external resources to reach decent performance (Otero, 2007). For example, D´ejean et al. (2002) com- bined a multilingual thesaurus and a dictionary to ex- tract bilingual lexicons from comparable corpora. Sim- ilarly, Nagata et al. (2001) relied on the web …

Content truncated to 500 words