Machine Translation System Combination by Confusion Forest

Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics, pages 1249–1257, Portland, Oregon, June 19-24, 2011. c⃝2011 Association for Computational Linguistics Machine Translation System Combination by Confusion Forest Taro Watanabe and Eiichiro Sumita National Institute of Information and Communications Technology 3-5 Hikaridai, Keihanna Science City, 619-0289 JAPAN {taro.watanabe,eiichiro.sumita}@nict.go.jp Abstract The state-of-the-art system combination method for machine translation (MT) is based on confusion networks constructed by aligning hypotheses with regard to word similarities. We introduce a novel system combination framework in which hypotheses are encoded as a confusion forest, a packed forest representing alternative trees. The forest is generated using syntactic consensus among parsed hypotheses: First, MT outputs are parsed. Second, a context free grammar is learned by extracting a set of rules that con- stitute the parse trees. Third, a packed forest is generated starting from the root symbol of the extracted grammar through non-terminal rewriting. The new hypothesis is produced by searching the best derivation in the forest. Experimental results on the WMT10 system combination shared task yield comparable performance to the conventional confusion network based method with smaller space. 1 Introduction System combination techniques take the advantages of consensus among multiple systems and have been widely used in fields, such as speech recognition (Fiscus, 1997; Mangu et al., 2000) or parsing (Hen- derson and Brill, 1999). One of the state-of-the-art system combination methods for MT is based on confusion networks, which are compact graph-based structures representing multiple hypotheses (Banga- lore et al., 2001). Confusion networks are constructed based on string similarity information. First, one skeleton or backbone sentence is selected. Then, other hypothe- ses are aligned against the skeleton, forming a lattice with each arc representing alternative word candi- dates. The alignment method is either model-based (Matusov et al., 2006; He et al., 2008) in which a statistical word aligner is used to compute hypothe- sis alignment, or edit-based (Jayaraman and Lavie, 2005; Sim et al., 2007) in which alignment is mea- sured by an evaluation metric, such as translation er- ror rate (TER) (Snover et al., 2006). The new trans- lation hypothesis is generated by selecting the best path through the network. We present a novel method for system combina- tion which exploits the syntactic similarity of system outputs. Instead of constructing a string-based con- fusion network, we generate a packed forest (Billot and Lang, 1989; Mi et al., 2008) which encodes ex- ponentially many parse trees in a polynomial space. The packed forest, or confusion forest, is constructed by merging the MT outputs with regard to their syntactic consensus. We employ a grammar-based method to generate the confusion forest: First, sys- tem outputs are parsed. Second, a set of rules are extracted from the parse trees. Third, a packed for- est is generated using a variant of Earley’s algorithm (Earley, 1970) starting from the unique root symbol. New hypotheses are selected by searching the best derivation in the forest. The grammar, a set of rules, is limited to those found in the parse …

Content truncated to 500 words