Enhancing Grammatical Cohesion: Generating Transitional Expressions for SMT

Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics, pages 850–860, Baltimore, Maryland, USA, June 23-25 2014. c⃝2014 Association for Computational Linguistics Enhancing Grammatical Cohesion: Generating Transitional Expressions for SMT Mei Tu Yu Zhou Chengqing Zong National Laboratory of Pattern Recognition, Institute of Automation, Chinese Academy of Sciences {mtu,yzhou,cqzong}@nlpr.ia.ac.cn Abstract Transitional expressions provide glue that holds ideas together in a text and enhance the logical organization, which together help im- prove readability of a text. However, in most current statistical machine translation (SMT) systems, the outputs of compound-complex sentences still lack proper transitional expres- sions. As a result, the translations are often hard to read and understand. To address this issue, we propose two novel models to en- courage generating such transitional expres- sions by introducing the source compound- complex sentence structure (CSS). Our models include a CSS-based translation model, which generates new CSS-based translation rules, and a generative transfer model, which en- courages producing transitional expressions during decoding. The two models are integrat- ed into a hierarchical phrase-based translation system to evaluate their effectiveness. The ex- perimental results show that significant im- provements are achieved on various test data meanwhile the translations are more cohesive and smooth. 1 Introduction During the last decade, great progress has been made on statistical machine translation (SMT) models. However, these translations still suffer from poor readability, especially translations of compound-complex sentences. One of the main reasons may be that most existing models con- centrate more on producing well-translated local sentence fragments, but largely ignore global cohesion between the fragments. Generally, co- hesion, including lexical and grammatical cohe- sion, contributes much to the understandability and smoothness of a text. Recently, researchers have begun addressing the lexical cohesion of SMT (Gong et al., 2011; Xiao et al., 2011; Wong and Kit, 2012; Xiong, 2013). These efforts focus mainly on the co- occurrence of lexical items in a similar environ- ment. Grammatical cohesion1 (Halliday and Has- san, 1976) in SMT has been little mentioned in previous work. Translations without grammatical cohesion is hard to read, mostly due to loss of cohesive and transitional expressions between two sentence fragments. Thus, generating transi- tional expressions is necessary for achieving grammatical cohesion. However, it is not easy to produce such transitional expressions in SMT. As an example, consider the Chinese-to-English translation in Figure 1. Source Chinese sentence: [尽管 减轻 污染 的 呼声 不断 ，]1 [ 公众 Although reduce pollution of calls continue , public 日渐 愤怒 ，]2 [污染 还是 变得 更 糟糕 growing angry , pollution still become more worse 了 ，]3 [越发 显出 环保 的 紧迫性 。]4 already , more show environment protection of urgent . Target English golden translation: Despite frequent calls for cutting pollution, and growing public anger, the problem has only got worse, which increasingly shows the urgency of environmental protection. Figure 1: An example of Chinese-to-English transla- tion. The English translation sentence has three transi- tional phrases: Despite, and, which. There are 4 sub-sentences separated by com- mas in the Chinese …

Content truncated to 500 words