The FBK Participation in the WMT15 Automatic Post-editing Shared Task

Proceedings of the Tenth Workshop on Statistical Machine Translation, pages 210–215, Lisboa, Portugal, 17-18 September 2015. c⃝2015 Association for Computational Linguistics. The FBK Participation in the WMT15 Automatic Post-editing Shared Task Rajen Chatterjee Fondazione Bruno Kessler chatterjee@fbk.eu Marco Turchi Fondazione Bruno Kessler turchi@fbk.eu Matteo Negri Fondazione Bruno Kessler negri@fbk.eu Abstract In this paper, we describe the “FBK English- Spanish Automatic Post-editing (APE)” systems submitted to the APE shared task at the WMT 2015. We explore the most widely used statistical APE technique (monolingual) and its most signif- icant variant (context-aware). In this exploration, we introduce some novel task-specific dense fea- tures through which we observe improvements over the default setup of these approaches. We show these features are useful to prune the phrase table in order to remove unreliable rules and help the decoder to select useful translation options during decoding. Our primary APE system sub- mitted at this shared task performs significantly better than the standard APE baseline. 1 Introduction Over the last decade a lot of research has been carried out to mimic the human post-editing pro- cess in the field of Automatic Post-Editing (APE). The objective of APE is to learn how to cor- rect machine translation (MT) errors leveraging the human post-editing feedback. The variety of data generated by human feedback, in terms of post editing, possess an unprecedented wealth of knowledge about the dynamics (practical and cog- nitive) of the translation process. APE leverages the potential of this knowledge to improve MT quality. The problem is appealing for several rea- sons. On one side, as shown by Parton et al. (2012), APE systems can improve MT output by exploiting information unavailable to the decoder, or by performing deeper text analysis that is too expensive at the decoding stage. On the other side, APE represents the only way to rectify er- rors present in the “black-box” scenario where the MT system is unknown or its internal decoding in- formation is not available. The goal of the APE task is to challenge the re- search groups to improve the MT output quality by the use of a dataset consisting of triplets of sen- tences (source, MT output, human post-edition). We are facing the “MT-as-Black-box” scenario, so neither we have access to the MT engine nor do we have any decoding trace. The data for this pi- lot task belongs to generic news domain which reflects data sparseness, and the post-edition of the MT output is obtained through crowdsourcing which makes it vulnerable to noise thus making this task even more challenging. To begin with, §2 discusses the statistical APE methods used to implement the APE systems. §3 describes the data set available for this shared task, and provides detail of the experimental setup. §4 is our major contribution which discusses the FBK-APE pipeline and shows that incorporation of task-specific dense features can be useful to en- hance APE systems. Our final submitted system is reported in §5 followed by conclusion in §6. 2 Statistical APE Methods In …

Content truncated to 500 words