Comparing Representations of Semantic Roles for String-To-Tree Decoding

Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1786–1791, October 25-29, 2014, Doha, Qatar. c⃝2014 Association for Computational Linguistics Comparing Representations of Semantic Roles for String-To-Tree Decoding Marzieh Bazrafshan and Daniel Gildea Department of Computer Science University of Rochester Rochester, NY 14627 Abstract We introduce new features for incorpo- rating semantic predicate-argument struc- tures in machine translation (MT). The methods focus on the completeness of the semantic structures of the translations, as well as the order of the translated seman- tic roles. We experiment with translation rules which contain the core arguments for the predicates in the source side of a MT system, and observe that using these rules significantly improves the translation quality. We also present a new semantic feature that resembles a language model. Our results show that the language model feature can also significantly improve MT results. 1 Introduction In recent years, there have been increasing ef- forts to incorporate semantics in statistical ma- chine translation (SMT), and the use of predicate- argument structures has provided promising im- provements in translation quality. Wu and Fung (2009) showed that shallow semantic parsing can improve the translation quality in a machine trans- lation system. They introduced a two step model, in which they used a semantic parser to rerank the translation hypotheses of a phrase-based sys- tem. Liu and Gildea (2010) used semantic fea- tures for a tree-to-string syntax based SMT sys- tem. Their features modeled deletion and reorder- ing for source side semantic roles, and they im- proved the translation quality. Xiong et al. (2012) incorporated the semantic structures into phrase- based SMT by adding syntactic and semantic fea- tures to their translation model. They proposed two discriminative models which included fea- tures for predicate translation and argument re- ordering from source to target side. Bazrafshan and Gildea (2013) used semantic structures in a string-to-tree translation system by extracting translation rules enriched with semantic informa- tion, and showed that this can improve the trans- lation quality. Li et al. (2013) used predicate- argument structure reordering models for hierar- chical phrase-based translation, and they used lin- guistically motivated constraints for phrase trans- lation. In this paper, we experiment with methods for incorporating semantics in a string-to-tree MT system. These methods are designed to model the order of translation, as well as the completeness of the semantic structures. We extract translation rules that include the complete semantic structure in the source side, and compare that with using semantic rules for the target side predicates. We present a method for modeling the order of seman- tic role sequences that appear spread across multi- ple syntax-based translation rules, in order to over- come the problem that a rule representing the en- tire semantic structure of a predicate is often too large and too specific to apply to new sentences during decoding. For this method, we compare the verb-specific roles of PropBank and the more gen- eral thematic roles of VerbNet. These essential arguments of a verbal …

Content truncated to 500 words