| With the massive growth of network information, summary extraction from mass documents becomes a hot topic in the Natural language processing area. For the past half century, researchers addressed problem for different points of view and proposed various solutions and paradigms. This paper discusses automatic summarization based on information extraction through calculating the sentence similarity with WordNet and Vector space model.Automatic summarization based on information extraction is to extract central contents from the given documents and generate summaries. Summaries are usually numbered, and we have to choose the highly-recapitulative sentences for summarization. We first extract five basic characteristics from the corpus, including sentence length, word frequency, sentence location, cue words and headline words. Then, each sentence is scored base on the five characteristics and sort all sentences in decreasing order of scores. Sentences with higher score are chosen as candidate ones for summary.This paper adopts semantic similarity to the automatic summarization, which can evaluate the word relevance. There are two manners to calculate sentence similarity:WordNet and Vector space model. Lexical similarity is obtained by Path and Lch. We deal with the polysemy words by two manners. One is to use the first meaning while the other is to use multiple meanings. Hence, we can evaluate the quality of summaries and analyze the key factor deciding the quality. For Vector space model, we also adopt two approaches:cosine and Dice. Throguh two different methods of calculating sentence similarity, we obtain a fuzzy similarity matrix, each elment of which represents the relevance between two sentences, and categorize the matrix by net and choose sentences for summarization. The experiment results show that introducing semantic similarity to automatic summarization can significantly improve the summury quality. |
