GHKM Rule Extraction and Scope-3 Parsing in Moses

Proceedings of the 7th Workshop on Statistical Machine Translation, pages 388–394, Montr´eal, Canada, June 7-8, 2012. c⃝2012 Association for Computational Linguistics GHKM Rule Extraction and Scope-3 Parsing in Moses Philip Williams and Philipp Koehn School of Informatics University of Edinburgh 10 Crichton Street EH8 9AB, UK p.j.williams-2@sms.ed.ac.uk pkoehn@inf.ed.ac.uk Abstract We developed a string-to-tree system for English–German, achieving competitive re- sults against a hierarchical model baseline. We provide details of our implementation of GHKM rule extraction and scope-3 parsing in the Moses toolkit. We compare systems trained on the same data using different gram- mar extraction methods. 1 Introduction Over the last few years, syntax-based rule extraction has largely developed along two lines, one originat- ing in hierarchical phrase-based translation (Chiang, 2005; Chiang, 2007) and the other in GHKM (Gal- ley et al., 2004; Galley et al., 2006). Hierarchical rule extraction generalizes the estab- lished phrase-based extraction method to produce formally-syntactic synchronous context-free gram- mar rules without any requirement for linguistic an- notation of the training data. In subsequent work, the approach has been extended to incorporate linguis- tic annotation on the target side (as in SAMT (Zoll- mann and Venugopal, 2006)) or on both sides (Chi- ang, 2010). In contrast, GHKM places target-side syntactic structure at the heart of the rule extraction process, producing extended tree transducer rules that map between strings and tree fragments. Ultimately, both methods define rules according to a sentence pair’s word-alignments. Without any restriction on rule size they will produce an expo- nentially large set of rules and so in practice only a subgrammar can be extracted. It is the differing rule selection heuristics that distinguish these two approaches, with hierarchical approaches being mo- tivated by phrasal coverage and GHKM by target- side tree coverage. The Moses toolkit (Koehn et al., 2007) has in- cluded support for hierarchical phrase-based rule ex- traction since the decoder was first extended to sup- port syntax-based translation (Hoang et al., 2009). In this paper we provide some implementation de- tails for the recently-added GHKM rule extractor and for the related scope-3 decoding algorithm. We then describe the University of Edinburgh’s GHKM- based English-German submission to the WMT translation task and present comparisons with hier- archical systems trained on the same data. To our knowledge, these are the first GHKM results pre- sented for English-German, a language pair with a high degree of reordering and rich target-side mor- phology. 2 GHKM Rule Extraction in Moses A basic GHKM rule extractor was first developed for Moses during the fourth Machine Translation Marathon1 in 2010. We have recently extended it to support several key features that are described in the literature, namely: composition of rules (Gal- ley et al., 2006), attachment of unaligned source words (Galley et al., 2004), and elimination of fully non-lexical unary rules (Chung et al., 2011). We provide some basic implementation details in the remainder of this section. In section 4 we present 1http://www.mtmarathon2010.info 388 TOP S-TOP PDS das VAFIN ist NP-PD ART der NN Fall PP-MNR APPR …

Content truncated to 500 words