A Trigger Language Model-based IR System

A Trigger Language Model-based IR System ZHANG Jun-lin SUN Le QU Wei-min SUN Yu-fang Open System & Chinese Information Processing Center Institute of Software, The Chinese Academy of Sciences P.O.BOX 8718,Beijing 100080 junlin01@iscas.cn Abstract Language model based IR system proposed in recent 5 years has introduced the language model approach in the speech recognition area into the IR community and improves the performance of the IR system effectively. However, the assumption that all the indexed words are irrelative behind the method is not the truth. Though statistical MT approach alleviates the situation by taking the synonymy factor into account, it never helps to judge the different meanings of the same word in varied context. In this paper we propose the trigger language model based IR system to resolve the problem. Firstly we compute the mutual information of the words from training corpus and then design the algorithm to get the triggered words of the query in order to fix down the topic of query more clearly. We introduce the relative parameters into the document language model to form the trigger language model based IR system. Experiments show that the performance of trigger language model based IR system has been improved greatly. The precision of trigger language model increased 12% and recall increased nearly 10.8% compared with Ponte language model method. 1 Introduction Using language models for information retrieval has been studied extensively recently(Jin et al 2002 Lafferty and Zhai 2001 Srikanth and Srihari 2002 Lavrenko and Croft 2001 Liu and Croft 2002). The basic idea is to compute the conditional probability P(Q|D), i.e. the probability of generating a query Q given the observation of a document D. Several different methods have been applied to compute this conditional probability. In most approaches, the computation is conceptually decomposed into two distinct steps: (1) Estimating a document language model; (2) Computing the query likelihood using the estimated document model based on some query model. For example, Ponte and Croft emphasized the first step, and used several heuristics to smooth the Maximum Likelihood of the document language model, and assumed that the query is generated under a multivariate Bernoulli model (Ponte and Croft 1998). The BBN method (Miller et al 1999) emphasized the second step and used a two-state hidden Markov model as the basis for generating queries, which, in effect, is to smooth the MLE with linear interpolation, a strategy also adopted in Hiemstra and Kraaij (Hiemstra and Kraaij 1999). In Zhai and Lafferty (Zhai and Lafferty 2001), it has been found that the retrieval performance is affected by both the estimation accuracy of document language models and the appropriate modeling of the query, and a two stage smoothing method was suggested to explicitly address these two distinct steps. It’s not hard to see that the unigram language model IR method contains the following assumption: Each word appearing in the document set and query has nothing to do with any other word. Obviously this assumption is not true in reality. Though statistical MT …

Content truncated to 500 words