Learning similarity functions in graph-based document summarization

Graph-based models have been extensively explored in document summarization in recent years. Compared with traditional feature-based models, graph-based models incorporate interrelated information into the ranking process. Thus, potentially they can do a better job in retrieving the important contents from documents. In this paper, we investigate the problem of how to measure sentence similarity which is a crucial issue in graph-based summarization models but in our belief has not been well defined in the past. We propose a supervised learning approach that brings together multiple similarity measures and makes use of human-generated summaries to guide the combination process. Therefore, it can be expected to provide more accurate estimation than a single cosine similarity measure. Experiments conducted on the DUC2005 and DUC2006 data sets show that the proposed learning approach is successful in measuring similarity. Its competitiveness and adaptability are also demonstrated.