Detecting Free Translation in Parallel Corpora from Attention Scores

Detecting Free Translation in Parallel Corpora from Attention Scores Qi Chen1,2 Oi Yee Kwong1∗ Jingbo Zhu2 1Department of Translation, The Chinese University of Hong Kong 2Natural Language Processing Lab, Northeastern University, China chenqi.neu@foxmail.com, oykwong@arts.cuhk.edu.hk, zhujingbo@mail.neu.edu.cn Abstract In this study, we propose a method for extract- ing free translation examples from bilingual parallel corpora based on an innovative use of attention scores. Preliminary results show that the approach is promising and paraphrases at both sentential and sub-sentential levels cover- ing diverse surface forms could be identified. The extracted data, upon further filtering, have great potential to supplement the example sen- tences available in existing bilingual dictionar- ies in an effective and systematic way. 1 Introduction In any language, the same meaning can often be expressed in alternative ways, or paraphrased. The recognition and generation of such meaning- equivalent forms are deemed important for var- ious natural language processing (NLP) applica- tions (Madnani and Dorr, 2010). In the inter-lingual context, typically in translation, there is inevitably a trace of paraphrasing, whether or not it is em- ployed consciously as a strategy in the process. Notwithstanding the different interpretation of ter- minology and research objectives across disciplines (e.g. translation vs paraphrase, literal translation vs free translation, etc. in translation studies, or para- phrase generation vs query expansion in the NLP community), access to such context-sensitive equiv- alents is essential especially when fluency, in addi- tion to fidelity, is concerned, for machine and hu- man translation alike. This study is thus motivated ∗ Corresponding author by the need to mine useful examples of free trans- lation for human translators’ reference on the one hand, and for inspiring machine translation’s further improvement on fluency on the other. Paraphrase, in this paper, is therefore used in a slightly restricted sense to refer to alternative expressions in a target language which are not only semantically equiva- lent, but also fulfill other contextual criteria to be qualified as translation of the text in a source lan- guage. Along the continuum of equivalence, free translation (as opposed to literal translation) will be of more interest to us. Three types of paraphrases are often included in paraphrase databases (e.g. Ganitkevitch et al., 2013; Ganitkevitch and Callison-Burch, 2014): lexical, phrasal, and syntactic paraphrases. Although the in- clusion of these paraphrase types has recognized that paraphrases could go beyond the replacement by synonyms or synonymous phrases, there is neverthe- less a restriction of syntactic category, as it at most allows the substitution of expressions that match a given syntactic type, even when non-constituent phrases are accommodated (e.g. Callison-Burch, 2008), as long as labels in syntactic trees are used as the point of reference. Paraphrases in different phrasal categories and syntactic constructions are particularly important in translation, as individual possible renditions would have their strengths and weaknesses in a given context with a certain literary style for a specific communicative purpose. In addition to fidelity, this is even more salient when fluency is concerned across language pairs with very distinct linguistic properties, where …

Content truncated to 500 words