UEdin: Translating L1 Phrases in L2 Context using Context-Sensitive SMT

Proceedings of the 8th International Workshop on Semantic Evaluation (SemEval 2014), pages 688–693, Dublin, Ireland, August 23-24, 2014. UEdin: Translating L1 Phrases in L2 Context using Context-Sensitive SMT Eva Hasler ILCC, School of Informatics University of Edinburgh e.hasler@ed.ac.uk Abstract We describe our systems for the SemEval 2014 Task 5: L2 writing assistant where a system has to find appropriate translations of L1 segments in a given L2 context. We participated in three out of four possible language pairs (English-Spanish, French- English and Dutch-English) and achieved the best performance for all our submit- ted systems according to word-based ac- curacy. Our models are based on phrase- based machine translation systems and combine topical context information and language model scoring. 1 Introduction In the past years, the fields of statistical machine translation (SMT) and word sense disambigua- tion (WSD) have developed largely in parallel, with each field organising their own shared tasks aimed at improving translation quality (Bojar et al., 2013) and predicting word senses, e.g. Agirre et al. (2010). Because sense disambiguation is a central problem in machine translation, there has been work on integrating WSD classifiers into MT systems (Carpuat and Wu, 2007a; Carpuat and Wu, 2007b; Chan et al., 2007). However, one problem with the direct integration of WSD techniques into MT has been the mismatch be- tween word predictions of WSD systems and the phrase segmentations of MT system. This prob- lem was adressed in Carpuat and Wu (2007b) by extending word sense disambiguation to phrase sense disambiguation. The relation between word sense distinctions and translation has also been explored in past SemEval tasks on cross-lingual word sense disambiguation, where senses are not This work is licensed under a Creative Commons Attribution 4.0 International Licence. Page numbers and proceedings footer are added by the organisers. Licence details: http: //creativecommons.org/licenses/by/4.0/ defined in terms of WordNet senses as in previ- ous tasks, but instead as translations to another language (Lefever and Hoste, 2010; Lefever and Hoste, 2013). This year’s L2 writing assistant task is simi- lar to the cross-lingual word sense disambigua- tion task but differs in the context provided for disambiguation and the length of the fragments (source phrases instead of words). While in other translation and disambiguation tasks the source language context is given, the L2 writing assis- tant task assumes a given target language con- text that constrains the possible translations of L1 fragments. This is interesting from a machine translation point-of-view because it allows for a direct comparison with systems that exploit the target context using a language model. As lan- guage models have become more and more power- ful over the years, mostly thanks to increased com- puting power, new machine translation techniques are also judged by their ability to improve perfor- mance over a baseline system with a strong lan- guage model. Another difference to previous Se- mEval tasks is the focus on both lexical and gram- matical forms, while previous tasks have mostly focused on lexical selection. 2 Translation Model for L1 Fragments …

Content truncated to 500 words