LIMSI@WMT’17

Proceedings of the Conference on Machine Translation (WMT), Volume 2: Shared Task Papers, pages 257–264 Copenhagen, Denmark, September 711, 2017. c⃝2017 Association for Computational Linguistics LIMSI@WMT’17 Franck Burlot and Pooyan Safari and Matthieu Labeau Alexandre Allauzen and Franc¸ois Yvon LIMSI, CNRS, Universit´e Paris Saclay, 91 403 Orsay, France firstname.lastname@limsi.fr Abstract This paper describes LIMSI’s submissions to the news shared task at WMT’17 for English into Czech and Latvian, as well as related experiments. This year’s novel- ties consist in the use of a neural machine translation system with a factored output predicting simultaneously a lemma deco- rated with morphological information and a fine-grained part-of-speech. Such a type of system drew our attention to the spe- cific step of reinflection, where lemmas and parts-of-speech are transformed into fully inflected words. Finally, we ran ex- periments showing an efficient strategy for parameter initialization, as well as data fil- tering procedures. 1 Introduction The contribution of LIMSI laboratory to the WMT 2017 News shared task consisted in the submis- sion of different systems for English-to-Czech, as well as with this year’s “guest” language pair: English-to-Latvian. Our main focus was on translation into mor- phologically rich languages (MRL), a challenging question in current state-of-the-art neural machine translation (NMT) architectures. Indeed, the va- riety of target word forms in these languages re- quires the use of an open vocabulary. To tackle this issue, we have experimented with a factored neu- ral machine translation system predicting simulta- neously at each timestep a normalized word and a fine-grained part-of-speech (section 3). A normal- ized word (section 5.2) is a specific representation where we removed part of the morphological con- tent of the word, keeping only the features that are relevant to the source language. Such a factored architecture required a non- trivial step consisting in reinflecting the MT pre- dictions, i.e. transforming normalized words and parts-of-speech into fully inflected words. To this end, we have experimented with a character-based language model that is used to select ambigu- ous word forms returned by a look-up table (sec- tion 5.5). Further experiments show the use of an auto- encoder to initialize the NMT system’s encoder (section 4.1), which enables a faster convergence of the parameter and therefore a lower training time. Finally, we report experiments performed with different data filtering procedures (section 4.2) and their impact on translation quality. 2 Data and Preprocessing The pre-processing of English data relies on in- house tools (D´echelotte et al., 2008). All the Czech data were tokenized and truecased using the Moses toolkit (Koehn et al., 2007). PoS-tagging was performed with Morphodita (Strakov´a et al., 2014). The pre-processing of Latvian was pro- vided by TILDE.1 Latvian PoS-tags were obtained with the LU MII Tagger (Paikens et al., 2013). All the data used to train our systems were provided at WMT’17.2 For English-to-Czech, the parallel data used consisted in nearly 20M sentences from a subset of WMT data relevant to the news domain: News- commentary, Europarl and specific categories of the Czeng corpus (news, …

Content truncated to 500 words