Utilization of Co-occurrence Relationships between Semantic Concepts in Re-ranking for Information Retrieval

Abstract

Semantic information retrieval is a popular research topic in the multimedia area. The goal of the retrieval is to provide the end users with as relevant results as possible. Many research efforts have been done to build ranking models for different semantic concepts (or classes). While some of them have been proven to be effective, others are still far from satisfactory. Our observation that certain target semantic concepts have high co-occurrence relationships with those easy-to-retrieve semantic concepts (or called reference semantics) has motivated us to utilize such co-occurrence relationships between semantic concepts in information retrieval and re-ranking. In this paper, we propose a novel semantic retrieval and re-ranking framework that takes advantage of the co-occurrence relationships between a target semantic concept and a reference semantic concept to re-rank the retrieved results. The proposed framework discretizes the training data into a set of feature-value pairs and employs Multiple Correspondence Analysis (MCA) to capture the correlation in terms of the impact weight between feature-value pairs and the positive-positive class in which the data instances belong to both the target semantic concept and the reference semantic concept. A combination of all these impact weights is utilized to re-rank the retrieved results for the target semantic concept. Comparative experiments are designed and evaluated on TRECVID 2005 and TRECVID 2010 video collections with public-available ranking scores. Experimental results on different retrieval scales demonstrate that our proposed framework can enhance the retrieval results for the target semantic concepts in terms of average precision, and the improvements for some semantic concepts are promising.