Translating Collocation using Monolingual and Parallel Corpus

Translating Collocation using Monolingual and Parallel Corpus ��� Ming-Zhuan Jiang, ��� Tzu-Xi Yen, ��� Chung-Chi Huang, ��� Mei-Hua Chen, ��� Jason S. Chang ����������� Dept of Computer Science National Tsing Hua Univ. {raconquer, joseph.yen, u901571, chen.meihua, jason.jschang}@gmail.com Abstract In this paper, we propose a method for translating a given verb-noun collocation based on a parallel corpus and an additional monolingual corpus. Our approach involves two models to generate collocation translations. The combination translation model generates combined translations of the collocate and the base word, and filters translations by a target language model from a monolingual corpus, and the bidirectional alignment translation model generates translations using bidirectional alignment information. At run time, each model generates a list of possible translation candidates, and translations in two candidate lists are re-ranked and returned as our system output. We describe the implementation of using method using Hong Kong Parallel Text. The experiment results show that our method improves the quality of top-ranked collocation translations, which could be used to assist ESL learners and bilingual dictionaries editors. Keyword: collocation, statistical machine translation, computer-assisted translation 1 INTRODUCTION A collocation is a recurrent combination of words that co-occur more frequently than expected by chance. Collocations can be classified into lexical and grammatical by the nature of their constituents. Another way of classifying collocations uses word positions to distinguish between rigid collocations and elastic collocations. Typically, a collocation consists of a base word and a collocate. Since collocations are used extensively, knowing the a right collocate for the base word plays an important role in second language learning as well as in machine translation. Translation of collocations is difficult for English as Second Language learners (ESL) because collocations are not always translated literally. For instance, the English collocation “delegate authority” can not be translated into “�� ��”. Much previous work has been done on collocation translation by extracting bilingual collocations pairs from parallel corpora. Recently, researchers have also proposed methods for retrieve collocations and their translation based on parsers and bilingual dictionaries. However, previous works using parallel corpora are mostly heuristic and methods based on bilingual dictionaries may be limited by the availability of broad-coverage dictionaries. Proceedings of the Twenty-Fourth Conference on Computational Linguistics and Speech Processing (ROCLING 2012) 246 More recently, the mainstream Statistical Machine Translation (SMT) system like Moses has been widely used in many translation tasks such as translating texts and sentences. Unfortunately, the traditional SMT system does not take into consideration of the structure of collocations including variable word forms and non-contiguous phrases. Little work has been done on improving the SMT system for finding flexible collocations translation as a tool to assist ESL learners or to help the task of compiling bilingual collocation dictionaries. Consider the elastic collocation “delegate ~ authority” and its translations. The translations of “authority” can be “��”, “��” and “���” which are found in parallel corpus. The traditional SMT system can find “delegate some authority” as “�� �� � �”, but usually there is no continuous “delegate authority” phrase …

Content truncated to 500 words