Query Lattice for Translation Retrieval

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 2031–2041, Dublin, Ireland, August 23-29 2014. Query Lattice for Translation Retrieval Meiping Dong†, Yong Cheng‡, Yang Liu†, Jia Xu‡, Maosong Sun†, Tatsuya Izuha⋄, Jie Hao# †State Key Laboratory of Intelligent Technology and Systems Tsinghua National Laboratory for Information Science and Technology Department of Computer Sci. and Tech., Tsinghua University, Beijing, China hellodmp@163.com, {liuyang2011,sms}@tsinghua.edu.cn ‡Institute for Interdisciplinary Information Sciences Tsinghua University, Beijing, China chengyong3001@gmail.com, xu@tsinghua.edu.cn ⋄Toshiba Corporation Corporate Research & Development Center tatsuya.izuha@toshiba.co.jp #Toshiba (China) R&D Center haojie@toshiba.com.cn Abstract Translation retrieval aims to find the most likely translation among a set of target-language strings for a given source-language string. Previous studies consider the single-best translation as a query for information retrieval, which may result in translation error propagation. To alleviate this problem, we propose to use the query lattice, which is a compact representation of exponentially many queries containing translation alternatives. We verified the effectiveness of query lattice through experiments, where our method explores a much larger search space (from 1 query to 1.24 × 1062 queries), runs much faster (from 0.75 to 0.13 second per sentence), and retrieves more accurately (from 83.76% to 93.16% in precision) than the standard method based on the query single-best. In addition, we show that query lattice significantly outperforms the method of (Munteanu and Marcu, 2005) on the task of parallel sentence mining from comparable corpora. 1 Introduction Translation retrieval aims to search for the most probable translation candidate from a set of target- language strings for a given source-language string. Early translation retrieval methods were widely used in example-based and memory-based translation systems (Sato and Nagao, 1990; Nirenburg et al., 1993; Baldwin and Tanaka, 2000; Baldwin, 2001). Often, the document set is a list of translation records that are pairs of source-language and target-language strings. Given an input source string, the retrieval system returns a translation record of maximum similarity to the input on the source side. Although these methods prove to be effective in example-based and memory-based translation systems, they heavily rely on parallel corpora that are limited both in size and domain. More recently, Liu et al. (2012) have proposed a new translation retrieval architecture that depends only on monolingual corpora. Given an input source string, their system retrieves translation candidates from a set of target-language sentences. This can be done by combining machine translation (MT) and information retrieval (IR): machine translation is used to transform the input source string to a coarse translation, which serves as a query to retrieve the most probable translation in the monolingual corpus. Therefore, it is possible for translation retrieval to have access to a huge volume of monolingual corpora that are readily available on the Web. However, the MT + IR pipeline suffers from the translation error propagation problem. Liu et al. (2012) use 1-best translations, which are inevitably erroneous due to the ambiguity and structural di- vergence of natural languages, as queries to the IR module. As a result, translation …

Content truncated to 500 words