Joint Parsing and Alignment with Weakly Synchronized Grammars

Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the ACL, pages 127–135, Los Angeles, California, June 2010. c⃝2010 Association for Computational Linguistics Joint Parsing and Alignment with Weakly Synchronized Grammars David Burkett John Blitzer Computer Science DivisionDan Klein University of California, Berkeley {dburkett,blitzer,klein}@cs.berkeley.edu Abstract Syntactic machine translation systems extract rules from bilingual, word-aligned, syntacti- cally parsed text, but current systems for pars- ing and word alignment are at best cascaded and at worst totally independent of one an- other. This work presents a unified joint model for simultaneous parsing and word alignment. To flexibly model syntactic divergence, we de- velop a discriminative log-linear model over two parse trees and an ITG derivation which is encouraged but not forced to synchronize with the parses. Our model gives absolute improvements of 3.3 F1 for English pars- ing, 2.1 F1 for Chinese parsing, and 5.5 F1 for word alignment over each task’s indepen- dent baseline, giving the best reported results for both Chinese-English word alignment and joint parsing on the parallel portion of the Chi- nese treebank. We also show an improvement of 1.2 BLEU in downstream MT evaluation over basic HMM alignments. 1 Introduction Current syntactic machine translation (MT) sys- tems build synchronous context free grammars from aligned syntactic fragments (Galley et al., 2004; Zollmann et al., 2006). Extracting such grammars requires that bilingual word alignments and mono- lingual syntactic parses be compatible. Because of this, much recent work in both word alignment and parsing has focused on changing aligners to make use of syntactic information (DeNero and Klein, 2007; May and Knight, 2007; Fossum et al., 2008) or changing parsers to make use of word align- ments (Smith and Smith, 2004; Burkett and Klein, 2008; Snyder et al., 2009). In the first case, how- ever, parsers do not exploit bilingual information. In the second, word alignment is performed with a model that does not exploit syntactic information. This work presents a single, joint model for parsing and word alignment that allows both pieces to influ- ence one another simultaneously. While building a joint model seems intuitive, there is no easy way to characterize how word align- ments and syntactic parses should relate to each other in general. In the ideal situation, each pair of sentences in a bilingual corpus could be syntacti- cally parsed using a synchronous context-free gram- mar. Of course, real translations are almost always at least partially syntactically divergent. Therefore, it is unreasonable to expect perfect matches of any kind between the two sides’ syntactic trees, much less expect that those matches be well explained at a word level. Indeed, it is sometimes the case that large pieces of a sentence pair are completely asyn- chronous and can only be explained monolingually. Our model exploits synchronization where pos- sible to perform more accurately on both word alignment and parsing, but also allows indepen- dent models to dictate pieces of parse trees and word alignments when synchronization is impossi- ble. This notion of “weak …

Content truncated to 500 words