LIMSI @ WMT’14 Medical Translation Task

Proceedings of the Ninth Workshop on Statistical Machine Translation, pages 246–253, Baltimore, Maryland USA, June 26–27, 2014. c⃝2014 Association for Computational Linguistics LIMSI @ WMT’14 Medical Translation Task Nicolas P´echeux1,2, Li Gong1,2, Quoc Khanh Do1,2, Benjamin Marie2,3, Yulia Ivanishcheva2,4, Alexandre Allauzen1,2, Thomas Lavergne1,2, Jan Niehues2, Aur´elien Max1,2, Franc¸ois Yvon2 Univ. Paris-Sud1, LIMSI-CNRS2 B.P. 133, 91403 Orsay, France Lingua et Machina3, Centre Cochrane franc¸ais4 {firstname.lastname}@limsi.fr Abstract This paper describes LIMSI’s submission to the first medical translation task at WMT’14. We report results for English- French on the subtask of sentence trans- lation from summaries of medical ar- ticles. Our main submission uses a combination of NCODE (n-gram-based) and MOSES (phrase-based) output and continuous-space language models used in a post-processing step for each system. Other characteristics of our submission in- clude: the use of sampling for building MOSES’ phrase table; the implementation of the vector space model proposed by Chen et al. (2013); adaptation of the POS- tagger used by NCODE to the medical do- main; and a report of error analysis based on the typology of Vilar et al. (2006). 1 Introduction This paper describes LIMSI’s submission to the first medical translation task at WMT’14. This task is characterized by high-quality input text and the availability of large amounts of training data from the same domain, yielding unusually high translation performance. This prompted us to experiment with two systems exploring differ- ent translation spaces, the n-gram-based NCODE (§2.1) and an on-the-fly variant of the phrase- based MOSES (§2.2), and to later combine their output. Further attempts at improving translation quality were made by resorting to continuous lan- guage model rescoring (§2.4), vector space sub- corpus adaptation (§2.3), and POS-tagging adap- tation to the medical domain (§3.3). We also per- formed a small-scale error analysis of the outputs of some of our systems (§5). 2 System Overview 2.1 NCODE NCODE implements the bilingual n-gram ap- proach to SMT (Casacuberta and Vidal, 2004; Mari˜no et al., 2006; Crego and Mari˜no, 2006) that is closely related to the standard phrase-based ap- proach (Zens et al., 2002). In this framework, the translation is divided into two steps. To translate a source sentence f into a target sentence e, the source sentence is first reordered according to a set of rewriting rules so as to reproduce the tar- get word order. This generates a word lattice con- taining the most promising source permutations, which is then translated. Since the translation step is monotonic, the peculiarity of this approach is to rely on the n-gram assumption to decompose the joint probability of a sentence pair in a sequence of bilingual units called tuples. The best translation is selected by maximizing a linear combination of feature functions using the following inference rule: e∗ = argmax e,a K � k=1 λkfk(f, e, a) (1) where K feature functions (fk) are weighted by a set of coefficients (λk) and a denotes the set of hidden variables corresponding to the reordering and segmentation of the source sentence. Along with the n-gram …

Content truncated to 500 words