Interactive-Predictive Machine Translation based on Syntactic Constraints of Prefix

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 1797–1806, Osaka, Japan, December 11-17 2016. Interactive-Predictive Machine Translation based on Syntactic Constraints of Prefix Na Ye, Guiping Zhang and Dongfeng Cai Human-Computer Intelligence Research Center Shenyang Aerospace University, Shenyang 110136, China yn.yena@hotmail.com, zgp@ge-soft.com, caidf@vip.163.com Abstract Interactive-predictive machine translation (IPMT) is a translation mode which combines ma- chine translation technology and human behaviours. In the IPMT system, the utilization of the prefix greatly affects the interaction efficiency. However, state-of-the-art methods filter trans- lation hypotheses mainly according to their matching results with the prefix on character level, and the advantage of the prefix is not fully developed. Focusing on this problem, this paper mines the deep constraints of prefix on syntactic level to improve the performance of IPMT systems. Two syntactic subtree matching rules based on phrase structure grammar are pro- posed to filter the translation hypotheses more strictly. Experimental results on LDC Chinese- English corpora show that the proposed method outperforms state-of-the-art phrase-based IPMT system while keeping comparable decoding speed. 1 Introduction In recent years, the machine translation (MT) technology has achieved great progress. However, up till now the MT output still cannot meet the practical requirements on translation quality in many scenari- os, and need to be post-edited by human translators before actually put to use. In order to help MT sys- tems collaborate with human translators more effectively, researchers carried out the study on comput- er-assisted translation (CAT), in which the main goal of MT is supporting professional translators in improving their productivity. Therefore, the ability to interact with human becomes an important re- search topic in CAT. The first interactive machine translation systems (Kay and Martins, 1973; Zajac, 1988; Yamron et al., 1993) focus on having human translators disambiguate the source texts through answering ques- tions. However, this question-answering process remains a laborious one for human translators. Under such circumstances, the interactive-predictive machine translation (IPMT) method is proposed (Foster et al., 1997). In the IPMT mode, first the system generates one or more raw suggestions, and then the human translator validates the longest correct prefix in the suggestions and revises the first character in the corresponding suffix, next the new prefix is used to help the system predict the optimal suffix. This process is repeated until the correct translation is acquired. The IPMT technology enables human translators to avoid the burden of explaining the source text, and directly control the final translation generation, so it attracted widespread attention. It can be seen that the essential difference between IPMT and MT is the introduction of a constraint, namely the target sentence must start with the human validated prefix. To achieve this goal, research- ers attempted various MT models (Och et al., 2003; Civera et al., 2004; Tomás and Casacuberta, 2006; Barrachina et al., 2009; González-Rubio et al., 2013). In these methods, prefix is mainly used for per- forming character-level matching on translation hypotheses to reduce the search space. However, the hypotheses that only match …

Content truncated to 500 words