Revising the Compositional Method for Terminology Acquisition from Comparable Corpora

Proceedings of COLING 2012: Technical Papers, pages 1797–1810, COLING 2012, Mumbai, December 2012. Revising the Compositional Method for Terminology Acquisition from Comparable Corpora Emmanuel Morin B´eatrice Daille Université de Nantes, LINA UMR CNRS 6241 2, rue de la Houssinière, BP 92208 F-44322 Nantes cedex 03 {emmanuel.morin,beatrice.daille}@univ-nant es.fr ABSTRACT In this paper, we present a new method that improves the alignment of equivalent terms monolingually acquired from bilingual comparable corpora: the Compositional Method with Context-Based Projection (CMCBP). Our overall objective is to identify and to translate high specialized terminology made up of multi-word terms acquired from comparable corpora. Our evaluation in the medical domain and for two pairs of languages demonstrates that CMCBP outperforms the state-of-art compositional approach commonly used for translationally equiv- alent multi-word term discovery from comparable corpora. KEYWORDS: Comparable corpora, bilingual lexicon, compositionality, multi-word term, con- text information. 1 Introduction The automatic compilation of bilingual dictionaries has received considerable attention in re- cent years for language for special purposes (LSP) (especially coming from scientific domains). LSP is characterised by the small amount of available textual data compared with general lan- guage, and a high proportion of specialised terms which are not be found in general language monolingual or bilingual dictionaries. For LSP, a specialised term could be either a single-word term (SWT) or a multi-word term (MWT), the latter being highly productive (Sag et al., 2002). A term is a lexical unit which represents a concept within a domain. As an example, in the medical domain, cancer is an SWT, breast cancer is an MWT. Comparable corpora that are sets of texts in two or more languages without being translations of each other, seem to be the right solution to solve the textual scarcity of LSP: as monolingual productions, they are authentic texts, and the babel web ensures that there is a sufficient num- ber of multilingual documents. The comparability of the corpus should be ensured by using various shared characteristics across languages that are checked during the compilation phase (McEnery, 2007). For LSP, the domain and sub-domain are requested as well as the commu- nicative settings and the textual genre to identify reliable translations (Bowker and Pearson, 2002). To build highly-specialised terminologies, the terms are first of all extracted monolingually from the comparable corpus. To collect close candidate terms across languages, it is necessary to use a term extraction program that applies the same method in the source and in the target languages. The translation of MWTs is the main need as they constitute around 80% of the domain-specific terms. See Nakagawa and Mori (2003) for the Japanese language. 1797 Our goal is to find the right translation for a source MWT in the set of MWT candidates in the target language. The simplest method assumes that the right translation of an MWT could be obtained by translating each component individually thanks to a general dictionary, by generat- ing all the combinations of word positions, and then filtering the translated expressions using either the list of target …

Content truncated to 500 words