Incremental Hypothesis Alignment with Flexible Matching for Building Confusion Networks: BBN System Description for WMT09 System Combination Task

Proceedings of the Fourth Workshop on Statistical Machine Translation , pages 61–65, Athens, Greece, 30 March – 31 March 2009. c⃝2009 Association for Computational Linguistics Incremental Hypothesis Alignment with Flexible Matching for Building Confusion Networks: BBN System Description for WMT09 System Combination Task Antti-Veikko I. Rosti and Bing Zhang and Spyros Matsoukas and Richard Schwartz BBN Technologies, 10 Moulton Street, Cambridge, MA 02138 � arosti,bzhang,smatsouk,schwartz � @bbn.com Abstract This paper describes the incremental hy- pothesis alignment algorithm used in the BBN submissions to the WMT09 system combination task. The alignment algo- rithm used a sentence specific alignment order, flexible matching, and new shift heuristics. These refinements yield more compact confusion networks compared to using the pair-wise or incremental TER alignment algorithms. This should reduce the number of spurious insertions in the system combination output and the sys- tem combination weight tuning converges faster. System combination experiments on the WMT09 test sets from five source languages to English are presented. The best BLEU scores were achieved by comb- ing the English outputs of three systems from all five source languages. 1 Introduction Machine translation (MT) systems have different strengths and weaknesses which can be exploited by system combination methods resulting in an output with a better performance than any indi- vidual MT system output as measured by auto- matic evaluation metrics. Confusion network de- coding has become the most popular approach to MT system combination. The first confusion net- work decoding method (Bangalore et al., 2001) was based on multiple string alignment (MSA) (Durbin et al., 1988) borrowed from biological sequence analysis. However, MSA does not al- low re-ordering. The translation edit rate (TER) (Snover et al., 2006) produces an alignment be- tween two strings and allows shifts of blocks of words. The availability of the TER software has made it easy to build a high performance system combination baseline (Rosti et al., 2007). The pair-wise TER alignment originally de- scribed by Sim et al. (2007) has various limita- tions. First, the hypotheses are aligned indepen- dently against the skeleton which determines the word order of the output. The same word from two different hypotheses may be inserted in differ- ent positions w.r.t. the skeleton and multiple inser- tions require special handling. Rosti et al. (2008) described an incremental TER alignment to miti- gate these problems. The incremental TER align- ment used a global order in which the hypotheses were aligned. Second, the TER software matches words with identical surface strings. The pair- wise alignment methods proposed by Ayan et al. (2008), He et al. (2008), and Matusov et al. (2006) are able to match also synonyms and words with identical stems. Third, the TER software uses a set of heuristics which is not always optimal in de- termining the block shifts. Karakos et al. (2008) proposed using inversion transduction grammars to produce different pair-wise alignments. This paper is organized as follows. A refined incremental alignment algorithm is described in Section 2. Experimental evaluation comparing the pair-wise and incremental TER alignment …

Content truncated to 500 words