Cross-lingual Pseudo Relevance Feedback Based on Weak Relevant Topic Alignment

Cross-lingual Pseudo Relevance Feedback Based on Weak Relevant Topic Alignment WANG Xu-wen Institute of Medical Information & Li- brary, Chinese Academy of Medical Sciences, Beijing 100020 wang.xuwen@imi- cams.ac.cn ZHANG Qiang State Grid Electric Power Research Insti- tute, Beijing 102200 zhangqiang7@sge pri.sgcc.com.cn WANG Xiao-jie Beijing University of Posts and Telecommu- nications, Beijing, 100876 xjwang@bupt.edu .cn LI Jun-lian Institute of Medical Information & Li- brary, Chinese Academy of Medical Sciences, Beijing 100020 li.junlian@imi- cams.ac.cn Abstract In this paper, a cross-lingual pseudo relevance feedback (PRF) model based on weak relevant topic alignment (WRTA) is proposed for cross language query expansion on unparallel web pages. Topics in different languages are aligned on the basis of translation. Useful ex- pansion terms are extracted from weak relevant topics according to the bilingual term similarity. Experiment results on web-derived unparalell data show the contribution of the WRTA-based PRF model to cross language information re- trieval. 1 Introduction The problem of word mismatch between queries and retrieved documents is particularly serious in cross language information retrieval (CLIR). The integration of query expansion techniques and translation knowledge is considered as an effective way to improve the CLIR performance (Ballesteros and Croft, 1998; Qu et al., 2000). Pseudo relevance feedback (PRF) is one of the useful query optimizing technologies for monolin- gual retrieval tasks (Rocchio, 1971; Ruthven and Lalmas, 2003). As to the CLIR task, researchers laid more efforts on establishing an effective cross-lin- gual PRF mechanism on the basis of the relevance and complementary of bilingual web pages (Balles- teros and Croft, 1997; Lavrenko et al., 2002). One of the key problems is how to choose useful or rel- evant bilingual expansion terms. Typical cross-lingual PRF methods assume the top retrieved documents are relevant and perform feedback calculations on the whole pseudo-relevant document level. High-frequency words are often used for expanding original queries. In recent years, topic models were applied to more and more multilingual tasks (Wang et al., 2009; Vulic et al., 2013). Ganguly (2012) proposed an im- proved cross-lingual topical relevance model based on the latent topics of top ranked documents. Wang (2013) proposed a cross-lingual PRF model based on bilingual topics and showed better results on par- allel or comparable corpus. However, the hypothe- sis of common shared bilingual topics is not always suitable for unparallel documents, since they are of- ten poor in content relevance. In most cases, web pages retrieved from differ- ent language fields for a specific query may lack of parallelism. There may be some common topics shared by the retrieved documents in both languages, but there are also some specific topics for source language retrieval results or target language re- trieval results respectively. Only the former com- mon shared topics would be helpful to cross-lingual PRF. PACLIC 29 529 29th Pacific Asia Conference on Language, Information and Computation pages 529 - 534 Shanghai, China, October 30 - November 1, 2015 Copyright 2015 by Xuwen Wang, Qiang Zhang, Xiaojie Wang and Junlian Li In this paper, we assume that retrieved results in …

Content truncated to 500 words