Combining Multi-Engine Machine Translation and Online Learning through Dynamic Phrase Tables

Combining Multi-Engine Machine Translation and Online Learning through Dynamic Phrase Tables Rico Sennrich Insitute of Computational Linguistics University of Zurich Binzm¨uhlestr. 14 CH-8050 Z¨urich sennrich@cl.uzh.ch Abstract Extending phrase-based Statistical Ma- chine Translation systems with a second, dynamic phrase table has been done for multiple purposes. Promising results have been reported for hybrid or multi-engine machine translation, i.e. building a phrase table from the knowledge of external MT systems, and for online learning. We argue that, in prior research, dynamic phrase ta- bles are not scored optimally because they may be of small size, which makes the Maximum Likelihood Estimation of trans- lation probabilities unreliable. We pro- pose basing the scores on frequencies from both the dynamic corpus and the primary corpus instead, and show that this modifi- cation significantly increases performance. We also explore the combination of multi- engine MT and online learning. 1 Introduction Two recent trends in Machine Translation are multi-engine MT, and online learning. In multi- engine MT, the aim is to combine the strengths of different MT systems to perform better than any single system. Online learning is of high interest in the field of interactive MT; In order to increase translation performance and user satisfaction, it is beneficial to consider previous post-edits made by the user of the system. Both approaches can be implemented within the phrase-based SMT framework by adding a second, dynamic phrase table. This architecture was first described in (Chen et al., 2007), who built a dy- namic phrase table trained on translation hypothe- c⃝ 2011 European Association for Machine Translation. ses by external MT systems. The online learn- ing system described in (Hardt and Elming, 2010) uses a similar architecture, with the difference that, rather than translations by external systems, previ- ous translations, post-edited by the user, constitute the dynamic corpus. The aim of this study is to: a) evaluate both ap- proaches in an independent reimplementation and on a different corpus; b) implement and evaluate an alternative scoring procedure that promises fur- ther performance gains; c) show the feasibility of combining multi-engine MT and online learning in a single framework. 2 Related Work System combination for Machine Translation is an active research field. The last two Workshops on Machine Translation (WMT) included a sys- tem combination task; an overview is given in (Callison-Burch et al., 2009; Callison-Burch et al., 2010). The effectiveness of system combination strongly depends on the relative performance of the systems being combined. In the 2009 WMT, (Callison-Burch et al., 2009) conclude that “In general, system combinations performed as well as the best individual systems, but not statistically significantly better than them.” A possible reason for this failure to improve on individual systems is given in the following year: “This year we excluded Google translations from the systems used in system combination. In last year’s evalua- tion, the large margin between Google and many of the other systems meant that it was hard to improve on when combining systems. This year, the system combinations perform better than …

Content truncated to 500 words