Cluster-based Model Fusion for Spontaneous Speech Retrieval

In this paper we present a new method for combining the results of different models in order to improve the performance on a difficult task: Information Retrieval from spontaneous speech. Our technique is based on clustering the training topics according to their tf-idf (term frequency-inverse document frequency) properties, and selecting the best models for each cluster. When the system runs on a test topic, the cluster of the topic needs to be determined and the combination of models of this cluster is used. We report significant improvement on the Malach test collection used at CLEF-CLSR 2007. We also include a comparison of the results of our method on automatic speech transcripts versus manual meta-data.