QCRI at WMT12: Experiments in Spanish-English and German-English Machine Translation of News Text

Proceedings of the 7th Workshop on Statistical Machine Translation, pages 298–303, Montr´eal, Canada, June 7-8, 2012. c⃝2012 Association for Computational Linguistics QCRI at WMT12: Experiments in Spanish-English and German-English Machine Translation of News Text Francisco Guzm´an, Preslav Nakov, Ahmed Thabet, Stephan Vogel Qatar Computing Research Institute Qatar Foundation Tornado Tower, floor 10, PO box 5825 Doha, Qatar {fguzman,pnakov,ahawad,svogel}@qf.org.qa Abstract We describe the systems developed by the team of the Qatar Computing Research Insti- tute for the WMT12 Shared Translation Task. We used a phrase-based statistical machine translation model with several non-standard settings, most notably tuning data selection and phrase table combination. The evaluation results show that we rank second in BLEU and TER for Spanish-English, and in the top tier for German-English. 1 Introduction The team of the Qatar Computing Research Insti- tute (QCRI) participated in the Shared Translation Task of WMT12 for two language pairs:1 Spanish- English and German-English. We used the state-of- the-art phrase-based model (Koehn et al., 2003) for statistical machine translation (SMT) with several non-standard settings, e.g., data selection and phrase table combination. The evaluation results show that we rank second in BLEU (Papineni et al., 2002) and TER (Snover et al., 2006) for Spanish-English, and in the top tier for German-English. In Section 2, we describe the parameters of our baseline system and the non-standard settings we experimented with. In Section 3, we discuss our primary and secondary submissions for the two lan- guage pairs. Finally, in Section 4, we provide a short summary. 1The WMT12 organizers invited systems translating be- tween English and four other European languages, in both di- rections: French, Spanish, German, and Czech. However, we only participated in Spanish→English and German→English. 2 System Description Below, in Section 2.1, we first describe our initial configuration; then, we discuss our incremental im- provements. We explored several non-standard set- tings and extensions and we evaluated their impact with respect to different baselines. These baselines are denoted in the tables below by a #number that corresponds to systems in Figures 1 for Spanish- English and in Figure 2 for German-English. We report case insensitive BLEU calculated on the news2011 testing data using the NIST scoring tool v.11b. 2.1 Initial Configuration Our baseline system can be summarized as follows: • Training: News Commentary + Europarl train- ing bi-texts; • Tuning: news2010; • Testing: news2011; • Tokenization: splitting words containing a dash, e.g., first-order becomes first @-@ order; • Maximum sentence length: 100 tokens; • Truecasing: convert sentence-initial words to their most frequent case in the training dataset; • Word alignments: directed IBM model 4 (Brown et al., 1993) alignments in both direc- tions, then grow-diag-final-and heuristics; • Maximum phrase length: 7 tokens; • Phrase table scores: forward & reverse phrase translation probabilities, forward & reverse lex- ical translation probabilities, phrase penalty; 298 • Language model: 5-gram, trained on the target side of the two training bi-texts; • Reordering: lexicalized, msd-bidirectional-fe; • Detokenization: reconnecting words that were split around dashes; • Model parameter optimization: minimum error rate …

Content truncated to 500 words