CNRC-TMT: Second Language Writing Assistant System Description

Proceedings of the 8th International Workshop on Semantic Evaluation (SemEval 2014), pages 192–197, Dublin, Ireland, August 23-24, 2014. CNRC-TMT: Second Language Writing Assistant System Description Cyril Goutte Michel Simard National Research Council Canada Multilingual Text Processing 1200 Montreal Road, Ottawa, Ontario K1A 0R6, Canada FirstName.LastName@nrc.ca Marine Carpuat Abstract We describe the system entered by the National Research Council Canada in the SemEval-2014 L2 writing assistant task. Our system relies on a standard Phrase-Based Statistical Machine Transla- tion trained on generic, publicly available data. Translations are produced by taking the already translated part of the sentence as fixed context. We show that translation systems can address the L2 writing assis- tant task, reaching out-of-five word-based accuracy above 80 percent for 3 out of 4 language pairs. We also present a brief analysis of remaining errors. 1 Introduction The Semeval L2 writing assistant task simulates the situation of an L2 language learner trying to translate a L1 fragment in a L2 context. This is clearly motivated by a L2 language learning sce- nario. However, a very similar scenario can be en- countered in Computer-Aided Translation. Trans- lation memories retrieve from a large corpus of al- ready translated documents the source segments that best match a new sentence to be translated. If an exact source match is found, the correspond- ing target translation can be expected to be suit- able with little or no post-editing. However, when only approximate matches are found, post-editing will typically be required to adapt the target side of the partially matching source segment to the source sentence under consideration. It is possible to automate this process: standard string matching algorithms and word alignment techniques can be used to locate the parts of the source segment that do not match the sentence to translate, and from c⃝2014, The Crown in Right of Canada. there the parts of the target segment that need to be modified (Bic¸ici and Dymetman, 2008; Simard and Isabelle, 2009; Koehn and Senellart, 2010). The task of translating a L1 fragment in L2 con- text therefore has much broader application than language learning. This motivation also provides a clear link of this task to the Machine Translation setting. There are also connections to the code- switching and mixed language translation prob- lems (Fung et al., 1999). In our work, we therefore investigate the use of a standard Phrase-Based Statistical Machine Translation (SMT) system to translate L1 frag- ments in L2 context. In the next section, we de- scribe the SMT system that we used in our submis- sion. We then describe the corpora used to train the SMT engine (Section 3), and our results on the trial and test data, as well as a short error analysis (Section 4). section 2 System Description The core Machine Translation engine used for all our submissions is Portage (Larkin et al., 2010), the NRC’s phrase-based SMT system. Given a suitably trained SMT system, the Task 5 input is processed as follows. For each sentence with an L1 fragment to …

Content truncated to 500 words