Experiments in Medical Translation Shared Task at WMT 2014

This paper describes Dublin City University’s (DCU) submission to the WMT 2014 Medical Summary task. We report our results on the test data set in the French to English translation direction. We also report statistics collected from the corpora used to train our translation system. We conducted our experiment on the Moses 1.0 phrase-based translation system framework. We performed a variety of experiments on translation models, reordering models, operation sequence model and language model. We also experimented with data selection and removal the length constraint for phrase-pair extraction. 1 System Description 1.1 Training Data Statistics and Preparation The training corpora provided to the medical translation shared task can be divided into 3 categories: Medical in-domain corpora: these corpora contain documents, patents, articles, terminology lists, and titles that are representative of the same medical domain as the development and test data sets (Table 1, second column). Medical out-of-domain corpora: these corpora also contain medical documents, patents, articles, terminologies lists and titles, but describe a different domain from the development and test data sets (Table 1, third column). General domain corpora: these corpora consist of general-domain text (WMT 2014 general translation subtask corpora), and encompass various domains. (We did not use these corpora in our system). Corpus In-domain Out-of-domain parallel sentence parallel sentence number number EMEA 1,092,568 0 COPPA 664,658 2,841,849 PatTR-title 408,502 2,096,270 PatTR-abstract 688,147 3,009,523 PatTR-claims 1,105,230 5,861,621