Efficient statistical machine translation with constrained reordering

Efficient Statistical Machine Translation with Constrained Reordering Evgeny Matusov, Stephan Kanthak, and Hermann Ney Lehrstuhl f¨ur Informatik VI, Computer Science Department, RWTH Aachen University D-52056 Aachen, Germany {matusov,kanthak,ney}@informatik.rwth-aachen.de. Abstract. This paper describes how word alignment information makes machine translation more effi- cient. Following a statistical approach based on finite-state transducers, we perform reordering of source sentences in training using automatic word alignments and estimate a phrase-based translation model. Using this model, we translate monotonically taking a permutation graph as input. The permutation graph is constrained using an efficient and flexible reordering framework. We then propose to automati- cally identify source word sequences which should always be translated monotonically and keep the word order of these sequences in search. This allows us to obtain fast good-quality translations. We present competitive experimental results on the Verbmobil German-to-English and BTEC Chinese-to-English translation tasks. 1 Introduction Word reordering is of crucial importance for ma- chine translation. Most of the phrase-based statisti- cal approaches like the Alignment Template system of (Och et al., 2004) rely on reorderings which are implicitly memorized with each pair of source and target phrases in training. Additional reorderings on phrase level are fully integrated into the decod- ing process, which increases the complexity of the system and makes it hard to modify. Other statistical approaches make use of the ef- ficient search representation with weighted finite- state transducers (WFSTs). Many of these ap- proaches use joint probabilities of the source and the target language string. The automated trans- ducer inference techniques OMEGA (Vilar, 2000) and GIATI (Casacuberta et al., 2004) estimate phrase-based models, but capture reordering only implicitly in bilingual corpus representations. This leads to a strong degradation of translation quality when translating into a language with a completely different word order. In (Bangalore et al., 2000) weighted reordering has been applied to target sen- tences. In order to reduce the computational com- plexity, this approach considers only a set of plau- sible reorderings seen on training data. In this paper, we follow a phrase-based joint- probability WFST translation approach, in which source sentence reordering is applied on word level, both in training and for translation. This is a novel approach inspired by the work of (Knight et al., 1998) and (Kumar et al., 2003). In this approach, a reordering graph is computed on-demand and taken as input for monotonic translation. The approach is modular and allows easy introduction of different reordering constraints and probabilistic dependen- cies. Here, we extend this approach by introduc- ing additional restrictions on reorderings. We de- scribe our efficient finite-state implementations of IBM (Berger et al., 1996), inverse IBM and local re- ordering constraints. Furthermore, we apply these constraints in the search, but keep the word or- der within source phrases which were consistently aligned monotonically in training. In the next section we review the general theory of our translation system based on weighted finite- state transducers and describe the use of word align- ments for reordering in training. We then discuss three modeling techniques …

Content truncated to 500 words