LIMSI Submission for WMT’17 Shared Task on Bandit Learning

Proceedings of the Conference on Machine Translation (WMT), Volume 2: Shared Task Papers, pages 674–679 Copenhagen, Denmark, September 711, 2017. c⃝2017 Association for Computational Linguistics LIMSI Submission for WMT’17 Shared Task on Bandit Learning Guillaume Wisniewski LIMSI, CNRS, Univ. Paris-Sud, Universit´e Paris-Saclay, 91 405 Orsay, France guillaume.wisniewski@limsi.fr Abstract This paper describes LIMSI participation to the WMT’17 shared task on Bandit Learning. The method we propose to adapt a seed system trained on out-domain data to a new, unknown domain relies on two components. First, we use a linear regression model to exploit the weak and partial feedback the system receives by learning to predict the reward a translation hypothesis will get. This model can then be used to score hypotheses in the search space and translate source sentences while taking into account the specificities of the in-domain data. Second, we use the UCB1 algorithm to choose which of the ‘adapted’ or ‘seed’ system must be used to translate a given source sentence in order to maxi- mize the cumulative reward. Results on the development and train sets show that the proposed method does not succeed in improving the seed system. We explore several hypotheses to explain this negative result. 1 Introduction The first Bandit Learning for Machine Translation shared task (Sokolov et al., 2017) aims at adapting a ‘seed’ MT system trained on out-domain corpora to a new domain considering only a ‘weak’ signal, namely a translation quality judgment rather than a reference translation or a post-edition. Such a situation arises when the user is not a skilled trans- lator but can nevertheless decide whether a trans- lation is useful or not. The signal is qualified as ‘weak’ as only the score of the translation pro- duced by a system can be known, the same sen- tence can not be translated twice and no reference is ever revealed. Adapting a MT system from a weak signal raises three main challenges. First, the parame- ters of the MT system must be estimated without knowing the reference translation which rules out most of the usual optimization methods for MT such as MERT, MIRA or the computation of like- lihood at the heart of NMT systems (Neubig and Watanabe, 2016). Second, the system must be trained in a ‘one-shot’ way as each source sen- tence can only be translated once and will result in a single reward. Third, no information about the target domain is available and its specificities must be discovered ‘on-the-fly’. To address these challenges, we propose an adaptation method that relies on two components. First, we use a linear regression model to exploit the weak and partial feedback the system receives by learning to predict the reward a translation hy- pothesis will get. This model can then be used to score hypotheses of the search space and trans- late source sentences while taking into account the specificities of the in-domain data. Second, we use the UCB1 algorithm to choose which of the ‘adapted’ or ‘seed’ system must be used to …

Content truncated to 500 words