Augmenting String-to-Tree and Tree-to-String Translation with Non-Syntactic Phrases

Proceedings of the Ninth Workshop on Statistical Machine Translation, pages 486–498, Baltimore, Maryland USA, June 26–27, 2014. c⃝2014 Association for Computational Linguistics Augmenting String-to-Tree and Tree-to-String Translation with Non-Syntactic Phrases Matthias Huck and Hieu Hoang and Philipp Koehn School of Informatics University of Edinburgh 10 Crichton Street Edinburgh EH8 9AB, UK {mhuck,hhoang,pkoehn}@inf.ed.ac.uk Abstract We present an effective technique to easily augment GHKM-style syntax-based ma- chine translation systems (Galley et al., 2006) with phrase pairs that do not comply with any syntactic well-formedness con- straints. Non-syntactic phrase pairs are distinguished from syntactic ones in or- der to avoid harming effects. We apply our technique in state-of-the-art string-to- tree and tree-to-string setups. For tree-to- string translation, we furthermore investi- gate novel approaches for translating with source-syntax GHKM rules in association with input tree constraints and input tree features. 1 Introduction Syntax-based statistical machine translation sys- tems utilize linguistic information that is obtained by parsing the training data. In tree-to-string translation, source-side syntactic tree annotation is employed, while string-to-tree translation exploits target-side syntax. The syntactic parse tree an- notation constrains phrase extraction to syntacti- cally well-formed phrase pairs: spans of syntactic phrases must match constituents in the parse tree. Standard phrase-based and hierarchical phrase- based statistical machine translation systems, in contrast, allow all phrase pairs that are consistent with the word alignment (Koehn et al., 2003; Chi- ang, 2005). A restriction of the phrase inventory to syntac- tically well-formed phrase pairs entails that possi- bly valuable information from the training data re- mains disregarded. While we would expect phrase pairs that are not linguistically motivated to be less reliable, discarding them altogether might be an overly harsh decision. The quality of an inventory of syntactic phrases depends heavily on the tree annotation scheme and the quality of the syntac- tic parses of the training data. Phrase pairs that do not span constituents in the tree annotation ob- tained from syntactic parses can provide reason- able alternative segmentations or alternative trans- lation options which prove to be valuable to the decoder. In this work, we augment the phrase invento- ries of string-to-tree and tree-to-string translation systems with phrase pairs that are not induced in the syntax-based extraction. We extract continu- ous phrases that are consistent with the word align- ment, without enforcing any constraints with re- spect to syntactic tree annotation. Non-syntactic phrases are added as rules to the baseline syntactic grammar with a fill-up technique. New rules are only added if their right-hand side does not exist yet. We extend the glue grammar with a special glue rule to allow for application of non-syntactic phrases during decoding. A feature in the log- linear model combination serves to distinguish non-syntactic phrases from syntactic ones. During decoding, the decoder can draw on both syntactic and non-syntactic phrase table entries and produce derivations which resort to both types of phrases. Such derivations yield hypotheses that make use of the alternative segmentations and translation op- tions provided through non-syntactic phrases. The search space is more diverse, and …

Content truncated to 500 words