The TALP-UPC Phrase-Based Translation Systems for WMT13: System Combination with Morphology Generation, Domain Adaptation and Corpus Filtering

Proceedings of the Eighth Workshop on Statistical Machine Translation, pages 134–140, Sofia, Bulgaria, August 8-9, 2013 c⃝2013 Association for Computational Linguistics The TALP-UPC Phrase-based Translation Systems for WMT13: System Combination with Morphology Generation, Domain Adaptation and Corpus Filtering Llu´ıs Formiga∗, Marta R. Costa-juss`a∗, Jos´e B. Mari˜no∗ Jos´e A. R. Fonollosa∗, Alberto Barr´on-Cede˜no∗†, Llu´ıs M`arquez∗ ∗TALP Research Centre †Facultad de Inform´atica Universitat Polit`ecnica de Catalunya Universidad Polit´ecnica de Madrid Barcelona, Spain Madrid, Spain {lluis.formiga,marta.ruiz,jose.marino,jose.fonollosa}@upc.edu {albarron, lluism}@lsi.upc.edu Abstract This paper describes the TALP participa- tion in the WMT13 evaluation campaign. Our participation is based on the combi- nation of several statistical machine trans- lation systems: based on standard phrase- based Moses systems. Variations include techniques such as morphology genera- tion, training sentence filtering, and do- main adaptation through unit derivation. The results show a coherent improvement on TER, METEOR, NIST, and BLEU scores when compared to our baseline sys- tem. 1 Introduction The TALP-UPC center (Center for Language and Speech Technologies and Applications at Univer- sitat Polit`ecnica de Catalunya) focused on the En- glish to Spanish translation of the WMT13 shared task. Our primary (contrastive) run is an internal system selection comprised of different train- ing approaches (without CommonCrawl, unless stated): (a) Moses Baseline (Koehn et al., 2007b), (b) Moses Baseline + Morphology Gener- ation (Formiga et al., 2012b), (c) Moses Baseline + News Adaptation (Henr´ıquez Q. et al., 2011), (d) Moses Baseline + News Adaptation + Mor- phology Generation , and (e) Moses Baseline + News Adaptation + Filtered CommonCrawl Adap- tation (Barr´on-Cede˜no et al., 2013). Our sec- ondary run includes is the full training strategy marked as (e) in the previous description. The main differences with respect to our last year’s participation (Formiga et al., 2012a) are: i) the inclusion of the CommonCrawl corpus, using a sentence filtering technique and the system com- bination itself, and ii) a system selection scheme to select the best translation among the different configurations. The paper is organized as follows. Section 2 presents the phrase-based system and the main pipeline of our baseline system. Section 3 de- scribes the our approaches to improve the baseline system on the English-to-Spanish task (special at- tention is given to the approaches that differ from last year). Section 4 presents the system combi- nation approach once the best candidate phrase of the different subsystems are selected. Section 5 discusses the obtained results considering both in- ternal and official test sets. Section 6 includes con- clusions and further work. 2 Baseline system: Phrase-Based SMT Our contribution is a follow up of our last year par- ticipation (Formiga et al., 2012a), based on a fac- tored Moses from English to Spanish words plus their Part-of-Speech (POS). Factored corpora aug- ments words with additional information, such as POS tags or lemmas. In that case, factors other than surface (e.g. POS) are usually less sparse, al- lowing the construction of factor-specific language models with higher-order n-grams. Such language models can help to obtain syntactically more cor- rect outputs. We used the standard …

Content truncated to 500 words