Shallow Semantically-Informed PBSMT and HPBSMT

Proceedings of the Eighth Workshop on Statistical Machine Translation, pages 177–184, Sofia, Bulgaria, August 8-9, 2013 c⃝2013 Association for Computational Linguistics Shallow Semantically-Informed PBSMT and HPBSMT Tsuyoshi Okita Qun Liu Josef van Genabith Dublin City University Glasnevin, Dublin 9, Ireland {tokita,qliu,josef}@computing.dcu.ie Abstract This paper describes shallow semantically-informed Hierarchical Phrase-based SMT (HPBSMT) and Phrase-Based SMT (PBSMT) systems developed at Dublin City University for participation in the translation task between EN-ES and ES-EN at the Work- shop on Statistical Machine Translation (WMT 13). The system uses PBSMT and HPBSMT decoders with multiple LMs, but will run only one decoding path decided before starting translation. Therefore the paper does not present a multi-engine system combination. We investigate three types of shallow seman- tics: (i) Quality Estimation (QE) score, (ii) genre ID, and (iii) context ID derived from context-dependent language models. Our results show that the improvement is 0.8 points absolute (BLEU) for EN-ES and 0.7 points for ES-EN compared to the standard PBSMT system (single best system). It is important to note that we developed this method when the standard (confusion network-based) system com- bination is ineffective such as in the case when the input is only two. 1 Introduction This paper describes shallow semantically- informed Hierarchical Phrase-based SMT (HPBSMT) and Phrase-Based SMT (PBSMT) systems developed at Dublin City University for participation in the translation task between EN-ES and ES-EN at WMT 13. Our objectives are to incorporate several shallow semantics into SMT systems. The first semantics is the QE score for a given input sentence which can be used to select the decoding path either of HPBSMT or PBSMT. Although we call this a QE score, this score is not quite a standard one which does not have access to translation output information. The second semantics is genre ID which is intended to capture domain adaptation. The third semantics is context ID: this context ID is used to adjust the context for the local words. Context ID is used in a continuous-space LM (Schwenk, 2007), but is implicit since the context does not appear in the construction of a continuous-space LM. Note that our usage of the term semantics refers to meaning constructed by a sentence or words. The QE score works as a sentence level switch to select HPBSMT or PBSMT, based on the semantics of a sentence. The genre ID gives an indication that the sentence is to be translated by genre ID- sensitive MT systems, again based on semantics on a sentence level. The context-dependent LM can be interpreted as supplying the local context to a word, capturing semantics on a word level. The architecture presented in this paper is sub- stantially different from multi-engine system com- bination. Although the system has multiple paths, only one path is chosen at decoding when process- ing unseen data. Note that standard multi-engine system combination using these three semantics has been presented before (Okita et al., 2012b; Okita et al., 2012a; Okita, 2012). This paper also compares the two approaches. The remainder …

Content truncated to 500 words