Learning from Post-Editing: Online Model Adaptation for Statistical Machine Translation

Proceedings of the 14th Conference of the European Chapter of the Association for Computational Linguistics, pages 395–404, Gothenburg, Sweden, April 26-30 2014. c⃝2014 Association for Computational Linguistics Learning from Post-Editing: Online Model Adaptation for Statistical Machine Translation Michael Denkowski Chris Dyer Alon Lavie Language Technologies Institute Carnegie Mellon University Pittsburgh, PA 15213 USA {mdenkows,cdyer,alavie}@cs.cmu.edu Abstract Using machine translation output as a starting point for human translation has become an increasingly common applica- tion of MT. We propose and evaluate three computationally efficient online methods for updating statistical MT systems in a scenario where post-edited MT output is constantly being returned to the system: (1) adding new rules to the translation model from the post-edited content, (2) updating a Bayesian language model of the target language that is used by the MT system, and (3) updating the MT system’s discriminative parameters with a MIRA step. Individually, these tech- niques can substantially improve MT qual- ity, even over strong baselines. Moreover, we see super-additive improvements when all three techniques are used in tandem. 1 Introduction Using machine translation outputs as a starting point for human translators is becoming increas- ingly common and is now arguably one of the most commercially important applications of MT. Con- siderable evidence has accumulated showing that human translators are more productive and accu- rate when post-editing MT output than when trans- lating from scratch (Guerberof, 2009; Carl et al., 2011; Koehn, 2012; Zhechev, 2012, inter alia). An important (if unsurprising) insight from prior research in this area is that translators become more productive as MT quality improves (Tat- sumi, 2009). While general improvements to MT continue to lead to further productivity gains, we explore how MT quality can be improved specifi- cally in an online post-editing scenario in which sentence-level MT outputs are constantly being presented to human experts, edited, and then re- turned to the system for immediate learning. This task is challenging in two regards. First, from a technical perspective, post-edited outputs must be processed rapidly: a productive post-editor cannot wait for a standard batch MT training pipeline to be rerun after each sentence is corrected! Sec- ond, from a methodological perspective, it is ex- pensive to run many human subject experiments, in particular when the human subjects must have translation expertise. We therefore use a sim- ulated post-editing paradigm in which either non-post-edited reference translations or manually post-edited translations from a similar MT system are used in lieu of human post-editors (§2). This paradigm allows us to efficiently develop and eval- uate systems that can go on to function in real-time post-editing scenarios without modification. We present and evaluate three online methods for improving translation models using feedback from editors: adding new translations rules to the translation grammar (§3), updating a Bayesian language model with observations of the post- edited output (§4), and using an online discrimi- native parameter update to minimize model error (§5). These techniques are computationally effi- cient and make minimal use of approximation or heuristics, handling initial and incremental data …

Content truncated to 500 words