Multi-document Summarization by Creating Synthetic Document Vector Based on Language Model

摘要

Multi-document summarization is to create summaries covering the major information that multiple documents tell in common. For this point, the existing methods are based on hand-crafted features for word and sentence. However, it is difficult to figure out the core contents of each document with the hand-crafted features because they have the limited information presented the given documents. Moreover, there exists a limit to figure out the major information because documents with the same meaning used to be paraphrased depending on their writers. Therefore, it is necessary to represent the semantic meanings of documents as well as sentences through understanding natural language. In this paper, we propose a new multi-document summarization system by creating a synthetic document vector covering the whole documents based on Language Model, whose is well-known for learning the semantic features in text. We experimented with DUC 2004 dataset provided by Document Understanding Conference (DUC) and the results show that our method summarizes multiple documents effectively based on their core contents.