MANY improvements for WMT’11

Proceedings of the 6th Workshop on Statistical Machine Translation, pages 135–139, Edinburgh, Scotland, UK, July 30–31, 2011. c⃝2011 Association for Computational Linguistics MANY improvements for WMT’11 Lo¨ıc Barrault LIUM, University of Le Mans Le Mans, France. FirstName.LastName@lium.univ-lemans.fr Abstract This paper describes the development op- erated into MANY for the 2011 WMT system combination evaluation campaign. Hypotheses from French/English and En- glish/French MT systems were combined with a new version of MANY, an open source system combination software based on confusion networks decoding currently developed at LIUM. MANY has been up- dated in order to optimize decoder pa- rameters with MERT, which proves to find better weights. The system combi- nation yielded significant improvements in BLEU score when applied on system com- bination data from two languages. 1 Introduction This year, the LIUM computer science laboratory participated in the French-English system combi- nation task at WMT’11 evaluation campaign. The system used for this task is MANY1 (Barrault, 2010), an open source system combination soft- ware based on Confusion Networks (CN). For this year evaluation, rather more technical than scientific improvements have been added to MANY. The tuning process has been improved by using MERT (Och, 2003) as a replacement of the numerical optimizer Condor (Berghen and Bersini, 2005). The impact of such change is de- tailed in section 3. After the evaluation period, some experiments have been performed on the English-French sys- tem combination task. The results are presented in the section 5. Before that, a quick description of MANY, including recent developments, can be found in section 2. 1MANY is available at the following address http:// www-lium.univ-lemans.fr/˜barrault/MANY 2 System description MANY is a system combination software (Bar- rault, 2010) based on the decoding of a lattice made of several Confusion Networks (CN). This is a widespread approach in MT system combina- tion (Rosti et al., 2007; Shen et al., 2008; Karakos et al., 2008; Rosti et al., 2009). MANY can be decomposed in two main modules. The first one is the alignment module which actually is a modi- fied version of TERp (Snover et al., 2009). Its role is to incrementally align the hypotheses against a backbone in order to create a confusion network. Those confusion networks are then connected to- gether to create a lattice. This module uses dif- ferent costs (which corresponds to a match, an in- sertion, a deletion, a substitution, a shift, a syn- onym and a stem) to compute the best alignment and incrementally build a confusion network. In the case of confusion network, the match (substi- tution, synonyms, and stems) costs are considered when the word in the hypothesis matches (is a sub- stitution, a synonyms or a stems of) at least one word of the considered confusion sets in the CN. System 0 System 1 TERp alignment LM output 1-best output 1-best output TERp alignment DECODE Merge System M 1-best output TERp alignment { best hypo nbest list Lattice CN CN CN Figure 1: System combination based on confusion network decoding. The second …

Content truncated to 500 words