Translation Divergence Analysis and Processing for Mandarin-English Parallel Text Exploitation

Translation Divergence Analysis and Processing for Mandarin-English Parallel Text Exploitation Shun-Chieh Lin, Jia-Ching Wang, and Jhing-Fa Wang Department of Electrical Engineering, National Cheng Kung University. 701 No.1, Ta-Hsueh Road, Tainan City Taiwan R.O.C. Tel: 886-6-2757575 Ext. 62341, Fax: 886-6-2761693 E-mail: wangjf@csie.ncku.edu.tw Abstract Previous work shows that the process of parallel text exploitation to extract transfer mappings between language pairs raises the capability of language translation. However, while this process can be fully automated, one thorny problem called “divergence” causes indisposed mapping extraction. Therefore, this paper discuss the issues of parallel text exploitation, in general, with special emphasis on divergence analysis and processing. In the experiments on a Mandarin-English travel conversation corpus of 11,885 sentence pairs, the perplexity with the alignments in IBM translation model is reduced averagely from 13.65 to 5.18 by sieving out inappropriate sentences from the collected corpus. 1. Introduction Over the past decade, research has focused on the automatic acquisition of translation knowledge from parallel text corpora. Statistical-based systems build alignment models from the corpora without linguistic analysis [1,2]. Another class of systems analyzes sentences in parallel texts to obtain transfer structures or rules [6]. Previous work shows that the process of parallel text exploitation to extract transfer mappings (models or rules) between language pairs can raise the capability of language translation. However, previous work is still hampered by the difficulties in transfer mapping extraction of achieving accurate lexical alignment and acquiring reusable structural correspondences. Although automatic extraction methods of lexical alignment and structural correspondences are introduced, they are not capable of handling exceptional cases like “divergence” presented in [4]. In general, divergence arises with variant lexical usage of role, position, and morphology between two languages. Therefore, while mapping extraction can be fully automated from parallel texts, divergence causes indisposed mapping extraction. Furthermore, the existence of translation divergences also makes adaptation from source structures into target structures difficult [5,7,8]. For parallel text exploitation, these divergences make the training process of transfer mapping extraction between languages impractical including parsing and word-level alignment, lexical-semantic lexicography, and syntactic structures. Therefore, study of parallel text exploitation needs a careful study of translation divergence. The framework of this paper is as follows. A brief overview of parallel text exploitation is discussed in Section 2. In Section 3, translation divergence analysis and processing for Mandarin-English parallel texts is presented. Section 4 shows experimental results with the alignments in IBM translation model. Finally, generalized conclusions are presented in Section 5. 2. Overview of Statistical-based Parallel Text Exploitation The goal of parallel text exploitation is to acquire the knowledge for translation of a text given in some source (“Mandarin”) string of words, m into a target (“English”) string of words, e. For the presented statistical approach [1] to string translation of ( Pre|m) , among all possible target strings, the string will be chosen with the highest probability which is given by Bayes’ decision rule as follows: |) argmaxPr()Pr( ˆ me e e e = (1) Pr(e) is the language model of target language and …

Content truncated to 500 words