Improved Translation with Source Syntax Labels

Proceedings of the Joint 5th Workshop on Statistical Machine Translation and MetricsMATR, pages 409–417, Uppsala, Sweden, 15-16 July 2010. c⃝2010 Association for Computational Linguistics Improved Translation with Source Syntax Labels Hieu Hoang School of Informatics University of Edinburgh h.hoang@sms.ed.ac.uk Philipp Koehn School of Informatics University of Edinburgh pkoehn@inf.ed.ac.uk Abstract We present a new translation model that include undecorated hierarchical-style phrase rules, decorated source-syntax rules, and partially decorated rules. Results show an increase in translation performance of up to 0.8% BLEU for German–English translation when trained on the news-commentary corpus, using syntactic annotation from a source lan- guage parser. We also experimented with annotation from shallow taggers and found this increased performance by 0.5% BLEU. 1 Introduction Hierarchical decoding is usually described as a formally syntactic model without linguistic com- mitments, in contrast with syntactic decoding which constrains rules and production with lin- guistically motivated labels. However, the decod- ing mechanism for both hierarchical and syntactic systems are identical and the rule extraction are similar. Hierarchical and syntax statistical machine translation have made great progress in the last few years and can claim to represent the state of the art in the field. Both use synchronous con- text free grammar (SCFG) formalism, consisting of rewrite rules which simultaneously parse the in- put sentence and generate the output sentence. The most common algorithm for decoding with SCFG is currently CKY+ with cube pruning works for both hierarchical and syntactic systems, as imple- mented in Hiero (Chiang, 2005), Joshua (Li et al., 2009), and Moses (Hoang et al., 2009) Rewrite rules in hierarchical systems have gen- eral applicability as their non-terminals are undec- orated, giving hierarchical system broad coverage. However, rules may be used in inappropriate sit- uations without the labeled constraints. The gen- eral applicability of undecorated rules create spu- rious ambiguity which decreases translation per- formance by causing the decoder to spend more time sifting through duplicate hypotheses. Syntac- tic systems makes use of linguistically motivated information to bias the search space at the expense of limiting model coverage. This paper presents work on combining hier- archical and syntax translation, utilizing the high coverage of hierarchical decoding and the in- sights that syntactic information can bring. We seek to balance the generality of using undeco- rated non-terminals with the specificity of labeled non-terminals. Specifically, we will use syntac- tic labels from a source language parser to label non-terminal in production rules. However, other source span information, such as chunk tags, can also be used. We investigate two methods for combining the hierarchical and syntactic approach. In the first method, syntactic translation rules are used con- currently with a hierarchical phrase rules. Each ruleset is trained independently and used concur- rently to decode sentences. However, results for this method do not improve. The second method uses one translation model containing both hierarchical and syntactic rules. Moreover, an individual rule can contain both decorated syntactic non-terminals, and undeco- rated hierarchical-style non-terminals (also, the left-hand-side non-terminal may, or may not be decorated). This results …

Content truncated to 500 words