Using Collocation Segmentation to Augment the Phrase Table

Proceedings of the Joint 5th Workshop on Statistical Machine Translation and MetricsMATR, pages 98–102, Uppsala, Sweden, 15-16 July 2010. c⃝2010 Association for Computational Linguistics Using collocation segmentation to augment the phrase table Carlos A. Henríquez Q.∗, Marta R. Costa-jussà†, Vidas Daudaravicius‡ Rafael E. Banchs†, José B. Mariño∗ ∗TALP Research Center, Universitat Politècnica de Catalunya, Barcelona, Spain {carlos.henriquez,jose.marino}@upc.edu †Barcelona Media Innovation Center, Barcelona, Spain {marta.ruiz,rafael.banchs}@barcelonamedia.org ‡Faculty of Informatics, Vytautas Magnus University, Kaunas, Lithuania vidas@donelaitis.vdu.lt Abstract This paper describes the 2010 phrase-based statistical machine translation system de- veloped at the TALP Research Center of the UPC1 in cooperation with BMIC2 and VMU3. In phrase-based SMT, the phrase table is the main tool in translation. It is created extracting phrases from an aligned parallel corpus and then computing trans- lation model scores with them. Performing a collocation segmentation over the source and target corpus before the alignment causes that dierent and larger phrases are extracted from the same original doc- uments. We performed this segmentation and used the union of this phrase set with the phrase set extracted from the non- segmented corpus to compute the phrase table. We present the con gurations con- sidered and also report results obtained with internal and o cial test sets. 1 Introduction The TALP Research Center of the UPC1 in coop- eration with BMIC2 and VMU3 participated in the Spanish-to-English WMT task. Our primary sub- mission was a phrase-based SMT system enhanced with POS tags and our contrastive submission was an augmented phrase-based system using colloca- tion segmentation (Costa-jussà et al., 2010), which mainly is a way of introducing new phrases in the translation table. This paper presents the descrip- tion of both systems together with the results that we obtained in the evaluation task and is organized as follows: rst, Section 2 and 3 present a brief de- scription of a phrase-based SMT, followed by a gen- eral explanation of collocation segmentation. Sec- tion 4 presents the experimental framework, corpus used and a description of the dierent systems built for the translation task; the section ends showing the results we obtained over the o cial test set. Fi- nally, section 5 presents the conclusions obtained from the experiments. 1Universitat Politècnica de Catalunya 2Barcelona Media Innovation Center 3Vytautas Magnus University 2 Phrase-based SMT This approach to SMT performs the translation splitting the source sentence in segments and as- signing to each segment a bilingual phrase from a phrase-table. Bilingual phrases are translation units that contain source words and target words, e.g. < unidad de traducci´on | translation unit >, and have dierent scores associated to them. These bilingual phrases are then sorted in order to max- imize a linear combination of feature functions. Such strategy is known as the log-linear model (Och and Ney, 2003) and it is formally de ned as: ˆe = arg max e � M � m=1 λmhm (e, f) � (1) where hm are dierent feature functions with weights λm. The two main feature functions are the translation model (TM) and …

Content truncated to 500 words