A Study of Document Weight Smoothness in Pseudo Relevance Feedback

In pseudo relevance feedback (PRF), the document weight which indicates how important a document is for the PRF model, plays a key role. In this paper, we investigate the smoothness issue of the document weights in PRF. The term smoothness means that the document weights decrease smoothly (i.e. gradually) along the document ranking list, and the weights are smooth (i.e. similar) within topically similar documents. We postulate that a reasonably smooth documentweighting function can benefit the PRF performance. This hypothesis is tested under a typical PRF model, namely the Relevance Model (RM). We propose a two-step document weight smoothing method, the different instantiations of which have different effects on weight smoothing. Experiments on three TREC collections show that the instantiated methods with better smoothing effects generally lead to better PRF performance. In addition, the proposed method can significantly improve the RM’s performance and outperform various alternative methods which can also be used to smooth the document weights.