Improving Abstractive Summarization with Graph Sequence Model

Based on dynamic data simulation, an article graph structure, namely a weighted adjacency matrix, is constructed to seek the similarity learning adjacency matrix corresponding to each pair of nodes in the document embedding space. In the process of generating summary results, the query embedding node matrix is used to convert text attributes into feature attributes. Then, the document structure is used to calculate forward and backward neighbours with edges as the main extension direction. The iteratively aggregated connected forward and backward representations are used as feature vector representations. Input the processed feature vector representation into a neural network model based on bidirectional attention mechanism and train it with other parts of the system to obtain a reference text summary summarizing the entire text. After comparing and evaluating the results, the method proposed in this article has improved the ROUGE score of the abstract to a certain extent and improved some of the problems such as word repetition and unclear sentences in the current text summarization process.