SUMSUG: Augmented Abstractive Text Summarization Model with Semantic Understanding Graphs

Abstract

As a research hotspot in natural language processing, automatic text summarization has been greatly developed. A summary should be a generalization based on depth understanding of the text. However, the existing models should not understand the semantic information of the text, so the generated summaries deviate from the semantics of the text and have low accuracy. This paper proposes an Augmented Abstractive Text Summarization Model with Semantic Understanding Graphs (SUMSUG). The model uses dual encoders, a text encoder and a graph encoder to guide the generation of summaries. And it obtains the features of the context from the text encoder and the structure features from the graph encoder. By fusing them the model obtains fuller semantic information. We evaluate the model on the Gigaword dataset. The experimental results show that the model performs better than other models, which prove the effectiveness of our model.