The UA-Prompsit hybrid machine translation system for the 2014 Workshop on Statistical Machine Translation

Proceedings of the Ninth Workshop on Statistical Machine Translation, pages 178–185, Baltimore, Maryland USA, June 26–27, 2014. c⃝2014 Association for Computational Linguistics The UA-Prompsit hybrid machine translation system for the 2014 Workshop on Statistical Machine Translation V´ıctor M. S´anchez-Cartagena,∗ ‡ Juan Antonio P´erez-Ortiz,∗ Felipe S´anchez-Mart´ınez∗ ∗Dep. de Llenguatges i Sistemes Inform`atics, Universitat d’Alacant, E-03071, Alacant, Spain ‡Prompsit Language Engineering, Av. Universitat, s/n. Edifici Quorum III, E-03202, Elx, Spain {vmsanchez,japerez,fsanchez}@dlsi.ua.es Abstract This paper describes the system jointly de- veloped by members of the Departament de Llenguatges i Sistemes Inform`atics at Universitat d’Alacant and the Promp- sit Language Engineering company for the shared translation task of the 2014 Workshop on Statistical Machine Trans- lation. We present a phrase-based sta- tistical machine translation system whose phrase table is enriched with information obtained from dictionaries and shallow- transfer rules like those used in rule-based machine translation. The novelty of our approach lies in the fact that the transfer rules used were not written by humans, but automatically inferred from a parallel cor- pus. 1 Introduction This paper describes the system jointly submitted by the Departament de Llenguatges i Sistemes In- form`atics at Universitat d’Alacant and the Promp- sit Language Engineering company to the shared translation task of the ACL 2014 Ninth Workshop on Statistical Machine Translation (WMT 2014). We participated in the English–French translation task with a hybrid system that combines, in a phrase-based statistical machine translation (PB- SMT) system, bilingual phrases obtained from par- allel corpora in the usual way (Koehn, 2010, ch. 5), and also bilingual phrases obtained from the existing dictionaries in the Apertium rule-based machine translation (RBMT) platform (Forcada et al., 2011) and a number of shallow-transfer ma- chine translation rules automatically inferred from a small subset of the training corpus. Among the different approaches for adding lin- guistic information to SMT systems (Costa-Juss`a and Farr´us, 2014), we followed the path we started with our submission to the Spanish–English WMT 2011 shared translation task (S´anchez-Cartagena et al., 2011b) which consisted of enriching the phrase table of a PBSMT system with phrase pairs generated using the dictionaries and rules in the Apertium (Forcada et al., 2011) Spanish–English RBMT system; our approach was one of the win- ners1 (together with two online SMT systems that were not submitted for the task but were included in the evaluation by the organisers and a system by Systran) in the pairwise manual evaluation of the English–Spanish translation task (Callison-Burch et al., 2011). In this submission, however, we only borrow the dictionaries from the Apertium English–French RBMT system and use them to au- tomatically infer the rules from a parallel corpus. We therefore avoid the need for human-written rules, which are usually written by trained experts, and explore a novel way to add morphological information to PBSMT. The rules inferred from corpora and used to enlarge the phrase table are shallow-transfer rules that build their output with the help of the bilingual dictionary and work on flat intermediate representations (see section 3.1); no syntactic parsing is …

Content truncated to 500 words