Online Learning Approaches in Computer Assisted Translation

Proceedings of the Eighth Workshop on Statistical Machine Translation, pages 301–308, Sofia, Bulgaria, August 8-9, 2013 c⃝2013 Association for Computational Linguistics Online Learning Approaches in Computer Assisted Translation Prashant Mathur‡†, Mauro Cettolo†, Marcello Federico† ‡ University of Trento † FBK - Fondazione Bruno Kessler Trento, Italy {prashant, cettolo, federico}@fbk.eu Abstract We present a novel online learning ap- proach for statistical machine translation tailored to the computer assisted transla- tion scenario. With the introduction of a simple online feature, we are able to adapt the translation model on the fly to the corrections made by the transla- tors. Additionally, we do online adaption of the feature weights with a large mar- gin algorithm. Our results show that our online adaptation technique outperforms the static phrase based statistical machine translation system by 6 BLEU points abso- lute, and a standard incremental adaptation approach by 2 BLEU points absolute. 1 Introduction The growing needs of the localization and trans- lation industry have recently boosted research around computer assisted translation (CAT) tech- nology. The purpose of CAT is to increase the pro- ductivity of a human translator. A CAT tool comes as a package of a Translation Memory (TM), built- in spell checkers, a dictionary, a terminology list etc. which help the translator while translating a sentence. Recent research has led to the in- tegration of CAT tools with statistical machine translation (SMT) engines. SMT makes use of a large available parallel corpus to generate statisti- cal models for translation. Due to their generaliza- tion capability, SMT systems are a good fit in this scenario and a seamless integration of SMT en- gines in CAT have shown to increase translator’s productivity (Federico et al., 2012). Although automatic systems generate reliable translations they are not accurate enough to be used directly and need postedition by human trans- lators. In state-of-the-art CAT tools, the SMT sys- tems are static in nature and so they cannot adapt to these corrections. When a SMT system keeps repeating the same error, productivity of transla- tors as well as their trust in SMT technology are negatively affected. As an example, technical doc- umentation typically contains a lot of repetitions due to the employed writing style and pervasive use of terminology. Hence, in order to provide useful hints, SMT systems are expected to behave consistently regarding the translation of domain- specific terms. However, if the user edits the trans- lation of a technical term in the target text, most current SMT systems are incapable to learn from those corrections. Online learning is a machine learning task where a predictor iteratively: (1) receives an input and outputs a label, (2) receives the correct label from a human and if the two labels do not match, it learns from the mistake. The task of learning from user corrections at the sentence level fits well the online learning scenario, and its expected useful- ness is clearly related to the amount of repetitions occurring in the text. The higher the number of repetititions in a …

Content truncated to 500 words