Ngram-based versus Phrase-based Statistical Machine Translation

Ngram-based versus Phrase-based Statistical Machine Translation Josep M. Crego, Marta R. Costa-juss`a, Jos´e B. Mari˜no and Jos´e A. R. Fonollosa TALP Research Center Universitat Polit`ecnica de Catalunya, Barcelona {jmcrego,mruiz,canton,adrian}@gps.tsc.upc.edu Abstract This work summarizes a comparison between two ap- proaches to Statistical Machine Translation (SMT), namely Ngram-based and Phrase-based SMT. In both approaches, the translation process is based on bilingual units related by word-to-word alignments (pairs of source and target words), while the main differences are based on the extraction process of these units and the sta- tistical modeling of the translation context. The study has been carried out on two different translation tasks (in terms of translation difficulty and amount of available training data), and allowing for distortion (reordering) in the decoding pro- cess. Thus it extends a previous work were both approaches were compared under monotone conditions. We finally report comparative results in terms of trans- lation accuracy, computation time and memory size. Re- sults show how the ngram-based approach outperforms the phrase-based approach by achieving similar accuracy scores in less computational time and with less memory needs. 1. Introduction From the initial word-based translation models [1], research on statistical machine translation has been strongly boosted. At the end of the last decade the use of context in the transla- tion model (phrase-based approach) lead to a clear improve- ment in translation quality ( [2], [3], [4]). Nowadays the in- troduction of some reordering abilities is of crucial impor- tance for some language pairs and is an important focus of research in the area of SMT. In parallel to the phrase-based approach, the ngram- based approach [5] also introduces the word context in the translation model, what allows to obtain comparable results under monotone conditions (as shown in [6]). The addi- tion of reordering abilities in the phrase-based approach is achieved by enabling a certain level of reordering in the source sentence. Though, the translation process consists of a composition of phrases, where the sequential composition of the phrases source words corresponds to the source sentence reordered. This procedure poses additional difficulties when applied to the ngram-based approach, because the character- istics of the ngram-based translation model. Despite of this, recent works ( [7], [8]) have shown how applying a reorder- ing schema in the training process the ngram-based approach can also take advantage of the distortion capabilities. In this paper we study the differences and similarities of both approaches (ngram-based and phrase-based), focusing on the translation model, where the translation context is dif- ferently taken into account. We also investigate the differ- ences in the translation (bilingual) units (tuples and phrases) and show efficiency results in terms of computation time and memory size for both systems. We have extended the com- parison in [6] to a Chinese to English task (where the use of distortion capabilities implies a clear improvement in trans- lation quality), and using a much larger Spanish to English task corpus. In section 2 we introduce the modeling underlying both SMT systems, the additional models …

Content truncated to 500 words