Subtopic-based Multi-documents Summarization

Abstract

Multi-documents summarization is an important research area of NLP. Most methods or techniques of multidocument summarization either consider the documents collection as single-topic or treat every sentence as single-topic only, but lack of a systematic analysis of the subtopic semantics hiding inside the documents. This paper presents a Subtopic-based Multi-documents Summarization (SubTMS) method. It adopts probabilistic topic model to discover the subtopic information inside every sentence and uses a suitable hierarchical subtopic structure to describe both the whole documents collection and all sentences in it. With the sentences represented as subtopic vectors, it assesses the semantic distances of sentences from the documents collection’s main subtopics and chooses sentences which have short distance as the final summary of the documents collection. In the experiments on DUC 2007 dataset, we have found that: when training a topic’s documents collection with some other topics’ documents collections as background knowledge, our approach can achieve fairly better ROUGE scores compared to other peer systems.