Generating Pseudo-relevant Representations for Spoken Document Retrieval

Spoken document retrieval (SDR) has become an important research subject due to the immenseness of multimedia data along with speech have spread around the world in our daily life. One of the fundamental challenge facing SDR is that the input query usually contains only a few words, which is too short to convey the information need of a user. In order to mitigate the problem, a well-practiced strategy is to reformulate the original query by performing a pseudo-relevance feedback process. Although several studies have evidenced its ability and capability for enhancing the retrieval performance, the time-consuming problem makes it hard to be used in reality. Motivated by the observations, in this paper, we concentrate on proposing a novel framework, which targets at generating a set of pseudo-relevant representations for a given query automatically, and eliminating the time-wasting problem. On top of the generated representations, we further investigate a novel query reformulation mechanism so as to improve the retrieval performance. A series of empirical SDR experiments conducted on a benchmark collection demonstrate the good efficacy of the proposed framework, compared to several existing strong baseline systems.