Using a Supertagged Dependency Language Model to Select a Good Translation in System Combination

Proceedings of NAACL-HLT 2013, pages 433–438, Atlanta, Georgia, 9–14 June 2013. c⃝2013 Association for Computational Linguistics Using a Supertagged Dependency Language Model to Select a Good Translation in System Combination Wei-Yun Ma Kathleen McKeown Department of Computer Science Department of Computer Science Columbia University Columbia University New York, NY 10027, USA New York, NY 10027, USA ma@cs.columbia.edu kathy@cs.columbia.edu Abstract We present a novel, structured language model - Supertagged Dependency Language Model to model the syntactic dependencies between words. The goal is to identify ungrammatical hypotheses from a set of candidate translations in a MT system combination framework and help select the best translation candidates using a variety of sentence-level features. We use a two-step mechanism based on constituent parsing and elementary tree extraction to obtain supertags and their dependency relations. Our experiments show that the structured language model provides significant improvement in the framework of sentence-level system combination. 1 Introduction In recent years, there has been a burgeoning interest in incorporating syntactic structure into Statistical machine translation (SMT) models (e..g, Galley et al., 2006; DeNeefe and Knight 2009; Quirk et al., 2005). In addition to modeling syntactic structure in the decoding process, a methodology for candidate translation selection has also emerged. This methodology first generates multiple candidate translations followed by rescoring using global sentence-level syntactic features to select the final translation. The advantage of this methodology is that it allows for easy integration of complex syntactic features that would be too expensive to use during the decoding process. The methodology is usually applied in two scenarios: one is as part of an n-best reranking (Och et al., 2004; Hasan et al., 2006), where n-best candidate translations are generated through a decoding process. The other is translation selection or reranking (Hildebrand and Vogel 2008; Callison-Burch et al., 2012), where candidate translations are generated by different decoding processes or different decoders. This paper belongs to the latter; the goal is to identify ungrammatical hypotheses from given candidate translations using grammatical knowledge in the target language that expresses syntactic dependencies between words. To achieve that, we propose a novel Structured Language Model (SLM) - Supertagged Dependency Language Model (SDLM) to model the syntactic dependencies between words. Supertag (Bangalore and Joshi, 1999) is an elementary syntactic structure based on Lexicalized Tree Adjoining Grammar (LTAG). Traditional supertagged n-gram LM predicts the next supertag based on the immediate words to the left with supertags, so it can not explicitly model long-distance dependency relations. In contrast, SDLM predicts the next supertag using the words with supertags on which it syntactically depend, and these words could be anywhere and arbitrarily far apart in a sentence. A candidate translation’s grammatical degree or “fluency” can be measured by simply calculating the SDLM likelihood of the supertagged dependency structure that spans the entire sentence. To obtain the supertagged dependency structure, the most intuitive way is through a LTAG parser (Schabes et al., 1988). However, this could be very 433 slow as it has time complexity of O(n6). Instead we propose an …

Content truncated to 500 words