Re-ranking Method Based on Topic Word Pairs

Re-ranking Method Based on Topic Word Pairs 1 Tingting He1 2 Ting Xu1 Guozhong Qu1 Xinhui Tu1 1 Department of Computer Science, Huazhong Normal University 430079 Wuhan, China 2 Software College of Tsinghua University 102201 Beijing, China tthe@mail.ccnu.edu.cn egg622@ mails.ccnu.edu.cn qu_g_z@mails.ccnu.edu.cn tuxinhui@mails.ccnu.edu.cn Abstract. How to improve the rankings of the relevant documents plays a key role in information retrieval. In this paper, a re-ranking approach based on topic words pair is proposed to improve precision while recall is preserved. The topic word pairs contain two correlated words, one of which is the original query word and the other come from the documents. The selection is based on Probabilistic Latent Semantic Indexing (PLSI). Then, the distribution of the word pairs is used to re-rank documents. Results show a 53.6% and 56.8% improvement compare to the initial retrieval without any re-ranking or query expansion on NTCIR-5 document collection for SLIR. 1 Introduction Since the number of electronic Chinese documents is growing very fast, efficient techniques for Chinese information retrieval are needed. In the past years, significant progress has been made on this problem by researchers in the field of information retrieval (IR) and many retrieval models, indexing strategies, query expansion and re-ranking strategies have been successfully used in IR. How to improve the rankings of the relevant documents plays a key role in information retrieval. In general, there are three kinds of information could be used for re-ranking: 1. Document information The clustering information has been used by Kyung-Soon Lee et al. [6]. Both the clusters and initial retrieval results have been used for re-ranking. Jaroslaw Balinski and Czeslaow Danilowicz [9] investigated ways to re-rank documents utilizing the inner distance between them. 2. Query information In the re-ranking phase, Mitra et al. [7] considered the correlation of query words. They add constraints to the query terms in order to get stronger indications for document relevance. Qu Youli et al. [8] described their method utilizing the location information of query terms. 3. Additional information Dequan Zheng et al.[10] adopt ontology information to re-rank documents. Currently, re-ranking using query information alone has some difficulties: this method neglects query words’ semantics. For example, document may be considered not relevant to query if they have not matching terms even though they are the same in semantics. Re-ranking using additional information (such as ontology, dictionary and so on) has other problem: the construction of ontology is still difficult. On the other hand, adding additional information into the query is equivalent to query expansion, which may cause a drift in the focus of the search topic. For example, 查询运动媒体或运动相关产业表彰老 虎伍兹为运动明星的报导(Find documents about sports media or related enterprises recognizing Tiger Woods as a sports star), many system will add “高尔夫”(golf) to the query, as a result documents will get high score if they contain “高尔夫” with high frequency, though the document talk about other linksman. Then in this paper, we present a re-ranking method utilizing both document information and query information. The main process is that: 1. Initial retrieval …

Content truncated to 500 words