Cascaded Phrase-Based Statistical Machine Translation Systems

Cascaded Phrase-Based Statistical Machine Translation Systems Dan Tufiş Research Institute for Artificial Intelligence, Romanian Academy Bucharest, Romania tufis@racai.ro Ștefan Daniel Dumitrescu Research Institute for Artificial Intelligence, Romanian Academy Bucharest, Romania sdumitrescu@racai.ro Abstract Statistical-based methods are the preva- lent approaches for implementing machine translation systems today. However the resulted translations are usually flawed to some degree. We assume that a statisti- cal baseline system can be re-used to automatically learn how to (partially) correct translation errors, i.e. to turn a “broken” target translation into a better one. By training and testing on initial bi- lingual data, we constructed a system S1 which was used to translate the source language part of the training corpus. The new translated corpus and its reference translation are used to train and test an- other similar system S2. Without any ad- ditional data, the chain S1+S2 shows a sensible quality increase against S1 in terms of BLEU scores, for both transla- tion directions (English to Romanian and Romanian to English). 1 Introduction The paper presents a cascaded phrase based translation system that obtains improved transla- tion scores using no additional data compared to the standard single-step translation system. The first challenge of our research was to ob- tain the best standard translation system possible. We experimented with different factored models that include surface form, lemmas and different part of speech tag sets in various combinations to confirm the assumption that translation accuracy is improved over a surface form only baseline model. The second objective of our work was to vali- date our intuition that a statistical baseline sys- tem can be re-used (cascaded) to automatically © 2012 European Association for Machine Translation. learn how to (partially) correct its own transla- tion errors, i.e. to turn an initially “broken” trans- lation into a better one. The phrase-based translation approach has overcome several drawbacks of the word-based translation methods and proved to significantly improve the quality of translated output. The morphology of a highly inflected language per- mits a flexible word order, thus shifting the focus from long-range reordering to the correct selec- tion of a morphological variant. Morphologically rich languages have a large number of surface forms in the lexicon to com- pensate for a flexible word order. Both Transfer and Interlingua MT employ a generation step to produce the surface form from a given context and a lemma of the word. In or- der to allow the same type of flexibility in using the morpho-syntactic information in translation, factored translation models (Koehn and Hoang, 2007) provide the possibility to integrate the lin- guistic information into the phrase-based transla- tion model. Most of the statistical machine translation (SMT) approaches that have a morphologically rich language as target employ factored transla- tion models. Our approach is similar to several other factored machine translation experiments such as adding the morphological features as fac- tors (Avramidis and Koehn, 2008). Our results confirm findings of other researchers, namely that when very large parallel corpora are avail- able, minimal pre-processing is sufficient …

Content truncated to 500 words