An efficient two-pass decoder for SMT using word confidence estimation

An Efficient Two-Pass Decoder for SMT Using Word Confidence Estimation Ngoc-Quang Luong Laurent Besacier LIG, Campus de Grenoble 41, Rue des Math´ematiques, UJF - BP53, F-38041 Grenoble Cedex 9, France {ngoc-quang.luong,laurent.besacier,benjamin.lecouteux}@imag.fr Benjamin Lecouteux Abstract During decoding, the Statistical Machine Translation (SMT) decoder travels over all complete paths on the Search Graph (SG), seeks those with cheapest costs and back- tracks to read off the best translations. Al- though these winners beat the rest in model scores, there is no certain guarantee that they have the highest quality with respect to the human references. This paper ex- ploits Word Confidence Estimation (WCE) scores in the second pass of decoding to enhance the Machine Translation (MT) quality. By using the confidence score of each word in the N-best list to update the cost of SG hypotheses containing it, we hope to “reinforce” or “weaken” them re- lied on word quality. After the update, new best translations are re-determined using updated costs. In the experiments on our real WCE scores and ideal (oracle) ones, the latter significantly boosts one-pass de- coder by 7.87 BLEU points, meanwhile the former yields an improvement of 1.49 points for the same metric. 1 Introduction Beside plenty of commendable achievements, the conventional one-pass SMT decoders are still not sufficient yet in yielding human-acceptable trans- lations (Zhang et al., 2006; Venugopal et al., 2007). Therefore, a number of methods to enhance them are proposed, such as: post-editing, re-ranking or re-decoding, etc. Post-editing (Parton et al., c⃝ 2014 The authors. This article is licensed under a Creative Commons 3.0 licence, no derivative works, attribution, CC- BY-ND. 2012) is in fact known to be a human-inspired task where the machine post edits translations in a second automatic pass. In re-ranking (Zhang et al., 2006; Duh and Kirchhoff, 2008; Bach et al., 2011), more features are integrated with the exist- ing multiple model scores for re-selecting the best candidate among N-best list. Meanwhile, the re- decoding process intervenes directly into the de- coder’s search graph (SG), driving it to the optimal path (cheapest hypothesis). The two-pass decoder has been built by several discrepant ways in the past. Kirchhoff and Yang (2005); Zhang et al. (2006) train additional Lan- guage Models (LM) and combine LM scores with existing model scores to re-rank the N-best list. Also focusing on the idea of re-ranking, yet Bach et al. (2011); Luong et al. (2014) employ sen- tence and word confidence scores in the second pass. Meanwhile, Venugopal et al. (2007) do a first pass translation without LM, but use it to score the pruned search hyper-graph in the second pass. This work concentrates on a second automatic pass where the costs of all hypotheses in the decoder’s SG containing words of the N-best list will be updated regarding the word quality predicted by Word Confidence Estimation (Ueffing and Ney, 2005) (WCE) system. In single-pass decoding, the decoder searches among complete paths (i.e. those cover all source words) for obtaining the optimal- cost ones. Essentially, the hypothesis …

Content truncated to 500 words