Interlocking Phrases in Phrase-based Statistical Machine Translation

Proceedings of NAACL-HLT 2016, pages 1076–1081, San Diego, California, June 12-17, 2016. c⃝2016 Association for Computational Linguistics Interlocking Phrases in Phrase-based Statistical Machine Translation Ye Kyaw Thu, Andrew Finch and Eiichiro Sumita Multilingual Translation Lab., Advanced Speech Translation Research and Development Promotion Center, 3-5 Hikaridai, Seika-cho, Soraku-gun, Kyoto, 619-0289, JAPAN {yekyawthu, andrew.finch, eiichiro.sumita}@nict.go.jp Abstract This paper presents an study of the use of in- terlocking phrases in phrase-based statistical machine translation. We examine the effect on translation quality when the translation units used in the translation hypotheses are allowed to overlap on the source side, on the target side and on both sides. A large-scale evaluation on 380 language pairs was conducted. Our results show that overall the use of overlap- ping phrases improved translation quality by 0.3 BLEU points on average. Further analysis revealed that language pairs requiring a larger amount of re-ordering benefited the most from our approach. When the evaluation was re- stricted to such pairs, the average improve- ment increased to up to 0.75 BLEU points with over 97% of the pairs improving. Our approach requires only a simple modification to the decoding algorithm and we believe it should be generally applicable to improve the performance of phrase-based decoders. 1 Introduction In this paper we examine the effect on machine translation quality of using interlocking phrases to during the decoding process in phrase-based statis- tical machine translation (PBSMT). The motivation for this is two-fold. Firstly, during the phrase-pair extraction process that occurs in the training of a typical PBSMT sys- tem, all possible alternative phrase-pairs are ex- tracted that are consistent with a set of alignment points. As a consequence, the source and tar- get sides of these extracted phrase pairs may over- lap. However, in contrast to this, the decoding pro- cess traditionally proceeds by concatenating disjoint translation units; the process relies on the language model to eliminate awkward hypotheses with re- peated words produced by sequences of translation units that overlap. Secondly, the transduction process in PBSMT is carried out by generating hypotheses that are com- posed of sequences of translation units. These se- quences are normally generated independently, as modeling the dependencies between them is difficult due to the data sparseness issues arising from model- ing with word sequences. The process of interlock- ing is a way of introducing a form of dependency be- tween translation units, effectively producing larger units from pairs of compatible units. 2 Related Work (Karimova et al., 2014) presented a method to ex- tract overlapping phrases offline for hierarchical phrase based SMT. They used the CDEC SMT de- coder (Dyer et al., 2010) that offers several learn- ers for discriminative tuning of weights for the new phrases. Their results showed improvements of 0.3 to 0.6 BLEU points over discriminatively trained hierarchical phrase-based SMT systems on two datasets for German-to-English translation. (Trib- ble and et al., 2003) proposed a method to gener- ate longer new phrases by merging existing phrase- level alignments that have overlaping words on both source …

Content truncated to 500 words