Improving Statistical Machine Translation with Selectional Preferences

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 2154–2163, Osaka, Japan, December 11-17 2016. Improving Statistical Machine Translation with Selectional Preferences Haiqing Tang, Deyi Xiong∗, Min Zhang and Zhengxian Gong Soochow University, Suzhou, China hqtang@stu.suda.edu.cn {dyxiong, minzhang, zhxgong}@suda.edu.cn Abstract Long-distance semantic dependencies are crucial for lexical choice in statistical machine trans- lation. In this paper, we study semantic dependencies between verbs and their arguments by modeling selectional preferences in the context of machine translation. We incorporate prefer- ences that verbs impose on subjects and objects into translation. In addition, bilingual selectional preferences between source-side verbs and target-side arguments are also investigated. Our ex- periments on Chinese-to-English translation tasks with large-scale training data demonstrate that statistical machine translation using verbal selectional preferences can achieve statistically sig- nificant improvements over a state-of-the-art baseline. 1 Introduction Lexical translation error is one of the most urgent issues for statistical machine translation (SMT). Al- though phrase-based SMT can deal with local context dependencies well, it performs rather poorly with long-distance dependencies and therefore causes a lot of lexical translation errors. Verbs and their ar- guments form such long-distance dependencies and play important roles in translation as they build skeletons of sentences. However, many SMT systems are not sufficient to capture long-distance depen- dencies between arguments and their dominating verbs. Verbs and arguments are often either incorrectly translated or not translated at all according to the error study by Wu and Fung (2009a). In order to address this issue, predicate-argument structures (PAS), which identify semantic frames within sentences by marking predicates, and labeling arguments with semantic roles, have been explored for SMT via various approaches in recent years. Wu and Fung (2009b) employ target-side PAS to pick out the most suitable translations among translation candidates after the decoding procedure is completed. Gildea (2010) integrates the PAS knowledge into decoding through projecting source-side PAS to the target-side via word alignments. In this paper, we are particularly interested in long-distance dependencies between verbs and their arguments in a predicate-argument structure. We propose to utilize selectional preferences (SPs) to handle these verb-argument dependencies for SMT. Selectional preferences place semantic restrictions on words, with which words can co-occur in dif- ferent syntactic patterns. To be more specific, the SPs of a verb can characterize the semantic restrictions that the verb imposes on its arguments. Violating these restrictions inevitably makes sentence senses odd or implausible. For example, in the sentence “The ball drinks a potato.”, both subject and object preferences for the verb “drink” are violated. SPs have proven useful for numerous applications, e.g., semantic role labeling (Gildea and Jurafsky, 2002), pronoun resolution (Bergsma et al., 2008), textual inference (Pantel et al., 2007), word-sense disambiguation (Resnik, 1997) and many more. Therefore, we have sufficient theoretical foundation to believe that SPs between verbs and arguments can be used to alleviate translation errors that we pointed out above. Our work consists of two parts: modeling SPs for verbs and incorporating SPs into an SMT system. In particular, …

Content truncated to 500 words