Generating document feature vector and extractive summary using sent2vec

In this paper, we propose a feature vector to express semantic contents of a document and proposes a method to generate the semantic summary of the document using the feature vector. First, a document is represented in a numerical matrix that is built up from a set of distributed representations of sentences, such as sent2vec. Next, singular value decomposition is applied to the matrix to generate three feature matrixes of the document. Then, we properly analyze a column vector, called singular vector, to choose feature vectors for expressing semantic contents of the document. The semantic summary is generated by each component of the feature vector, or by calculating distances between semantic content vector and sentences in the document. As a result of experiment, we found the feature vector represented the main and sub topics of the document. Furthermore, as a result of the experiment with a dataset of text summarization called DUC2002, we found the generating semantic summary shared 45% of the words in single document summarization and 38% of the words in multiple document summarization against the manually created summary