A new Re-ranking method for generic Chinese text summarization and its evaluation

Abstract

In this paper a new EMD-MMR (EMD: earth mover's distance; MMR: maximal marginal relevance) re-ranking method is proposed for generic Chinese text summarization. Our extraction-based summarization approach first ranks the sentences in a document by their weight calculated based on word frequency and position, and then re-ranks a few highly weighted sentences by the EMD-MMR method for sentence extraction. The proposed re-ranking method adopts a novel EMD-based similarity metric instead of the Cosine metric into the MMR approach. The EMD-based similarity metric can naturally take into account the semantic relatedness between words and compute the semantic similarity between texts with a many-to-many matching among words. We evaluate the performance of the proposed approach with a novel nk-blind method and the results demonstrate its effectiveness. © Springer-Verlag Berlin Heidelberg 2005.