Multi-source Neural Automatic Post-Editing: FBK’s participation in the WMT 2017 APE shared task

Proceedings of the Conference on Machine Translation (WMT), Volume 2: Shared Task Papers, pages 630–638 Copenhagen, Denmark, September 711, 2017. c⃝2017 Association for Computational Linguistics Multi-source Neural Automatic Post-Editing: FBK’s participation in the WMT 2017 APE shared task Rajen Chatterjee1,2, Amin Farajian1,2, Matteo Negri2, Marco Turchi2, Ankit Srivastava3, Santanu Pal4 1University of Trento 2Fondazione Bruno Kessler 3 German Research Center for Artificial Intelligence 4Saarland University Abstract Previous phrase-based approaches to Au- tomatic Post-editing (APE) have shown that the dependency of MT errors from the source sentence can be exploited by jointly learning from source and target informa- tion. By integrating this notion in a neu- ral approach to the problem, we present the multi-source neural machine transla- tion (NMT) system submitted by FBK to the WMT 2017 APE shared task. Our system implements multi-source NMT in a weighted ensemble of 8 models. The n-best hypotheses produced by this en- semble are further re-ranked using fea- tures based on the edit distance between the original MT output and each APE hy- pothesis, as well as other statistical models (n-gram language model and operation se- quence model). This solution resulted in the best system submission for this round of the APE shared task for both en-de and de-en language directions. For the for- mer language direction, our primary sub- mission improves over the MT baseline up to -4.9 TER and +7.6 BLEU points. For the latter, where the higher quality of the original MT output reduces the room for improvement, the gains are lower but still significant (-0.25 TER and +0.3 BLEU). 1 Introduction Automatic post-editing (APE) aims to correct sys- tematic machine translation (MT) errors, thereby reducing translators workload and eventually in- creasing translation productivity. The task, well motivated in (Bojar et al., 2015) and (Bojar et al., 2016), becomes necessary when working in a “black-box” condition where the MT engine used to translate is not directly accessible for retrain- ing or for more radical internal modifications. As pointed out in (Bojar et al., 2015), from the appli- cation point of view an APE system can help to: i) improve MT output by exploiting information un- available to the decoder, or by performing deeper text analysis that is too expensive at the decoding stage; ii) provide professional translators with im- proved MT output quality to reduce (human) post- editing effort and iii) adapt the output of a general- purpose MT system to the lexicon/style requested in a specific application domain. Different APE paradigms based on statistical methods (Simard et al., 2007; Dugast et al., 2007; Isabelle et al., 2007; Lagarda et al., 2009; Potet et al., 2012; Rosa et al., 2013; Lagarda et al., 2015; Chatterjee et al., 2017) have been proposed in the past showing the effectiveness of APE sys- tems. In the previous round of the APE shared task (WMT16), neural (Junczys-Dowmunt and Grund- kiewicz, 2016), hybrid (Chatterjee et al., 2016), and phrase-based (Pal et al., 2016b) solutions were all able to significantly improve MT output qual- ity in …

Content truncated to 500 words