Pointer-Generator Abstractive Text Summarization Model with Part of Speech Features

Abstract

The typical sequence-to-sequence with attention mechanism models have achieved good results in the task of abstractive text summarization. However, this kind of models always have some shortcomings: they have out-of-vocabulary (OOV) problems, sometimes may repeat themselves and are always of low quality. In order to solve these problems, we propose a pointer-generator text summarization model with part of speech features. First, we use the word vector and prat of speech vector as the input of the model, and then improve the quality of generated abstracts by combining convolutional neural network (CNN) and bi-directional LSTM. Second, we use pointergenerator network to control whether generating or copying words to solve the problem of OOV. Finally, we use coverage mechanism to monitor the abstract we have generated to avoid duplication problems. Compared with the classic pointergenerator network, the ROUGE scores of our model have greatly improved and the performance on LCSTS dataset is better than the state-of-the-art model at present.