The Study of Methods for Language Model Based Positive and Negative Relevance Feedback in Information Retrieval

Abstract

Relevance feedback techniques are important to Information retrieval (IR), which can effectively improve the performance of IR. The feedback includes positive and negative relevance one. The most of the previous work using feedback have focused on positive relevance feedback and pseudo relevance feedback in IR. In recent years, some work has been done and investigated the negative relevance feedback in IR. However, this paper highlights the incorporation or integration between the language models based positive and negative relevance feedback in IR, and through positive and negative feedback documents proportion on queries classification, with different parameters adjustment of positive and negative feedback ratio, where both types of feedback are used to modify and expand the user's query model. Our experimental results of using several TREC collections show that this method is significantly outperform the relevance feedback and pseudo relevance feedback in terms of the retrieval accuracy.