Abstract:Graph-based semi-supervised learning has proven to be an effective approach for query-focused multi-document summarization. The problem of previous semi-supervised learning is that sentences are ranked without considering the higher level information beyond sentence level. Researches on general summarization illustrated that the addition of topic level can effectively improve the summary quality. Inspired by previous researches, we propose a two-layer (i.e. sentence layer and topic layer) graph-based semi-supervised learning approach. At the same time, we propose a novel topic model which makes full use of the dependence between sentences and words. Experimental results on DUC and TAC data sets demonstrate the effectiveness of our proposed approach.

Query-focused Multi-document Summarization: Combining a Novel Topic Model with Graph-based Semi-supervised Learning