Analysis of Statistical and Morphological Classes to Generate Weigthed Reordering Hypotheses on a Statistical Machine Translation System

Proceedings of the Second Workshop on Statistical Machine Translation, pages 171–176, Prague, June 2007. c⃝2007 Association for Computational Linguistics Analysis of statistical and morphological classes to generate weighted reordering hypotheses on a Statistical Machine Translation system Marta R. Costa-juss`a and Jos´e A. R. Fonollosa Department of Signal Theory and Communications TALP Research Center (UPC) Barcelona 08034, Spain (mruiz,adrian)@gps.tsc.upc.edu Abstract One main challenge of statistical machine trans- lation (SMT) is dealing with word order. The main idea of the statistical machine reordering (SMR) approach is to use the powerful tech- niques of SMT systems to generate a weighted reordering graph for SMT systems. This tech- nique supplies reordering constraints to an SMT system, using statistical criteria. In this paper, we experiment with different graph pruning which guarantees the translation quality improvement due to reordering at a very low in- crease of computational cost. The SMR approach is capable of generalizing re- orderings, which have been learned during train- ing, by using word classes instead of words themselves. We experiment with statistical and morphological classes in order to choose those which capture the most probable reorderings. Satisfactory results are reported in the WMT07 Es/En task. Our system outperforms in terms of BLEU the WMT07 Official baseline system. 1 Introduction Nowadays, statistical machine translation is mainly based on phrases (Koehn et al., 2003). In parallel to this phrase- based approach, the use of bilingual n-grams gives com- parable results, as shown by Crego et al. (2005). Two basic issues differentiate the n-gram-based system from the phrase-based: training data is monotonically seg- mented into bilingual units; and, the model considers n- gram probabilities rather than relative frequencies. The n-gram-based system follows a maximum entropy ap- proach, in which a log-linear combination of multiple models is implemented (Mari˜no et al., 2006), as an al- ternative to the source-channel approach. Introducing reordering capabilities is important in both systems. Recently, new reordering strategies have been proposed such as the reordering of each source sentence to match the word order in the corresponding target sen- tence, see Kanthak et al. (2005) and Mari˜no et al. (2006). These approaches are applied in the training set and they lack of reordering generalization. Applied both in the training and decoding step, Collins et al. (2005) describe a method for introducing syntac- tic information for reordering in SMT. This approach is applied as a pre-processing step. Differently, Crego et al. (2006) presents a reordering approach based on reordering patterns which is coupled with decoding. The reordering patterns are learned di- rectly from word alignment and all reorderings have the same probability. In our previous work (Costa-juss`a and Fonollosa, 2006) we presented the SMR approach which is based on using the powerful SMT techniques to generate a re- ordered source input for an SMT system both in train- ing and decoding steps. One step further, (Costa-juss`a et al., 2007) shows how the SMR system can generate a weighted reordering graph, allowing the SMT system to make the final reordering decision. In this paper, …

Content truncated to 500 words