A global and local information extraction model incorporating selection mechanism for abstractive text summarization

A global and local information extraction model incorporating selection mechanism is presented to address the concerns of insufficient semantic coding and redundant semantic information in the abstractive summary. The model, unlike single coding, encodes the source text twice. The global semantic information is extracted by the encoder based on Bidirectional Gated Recurrent Unit network, while the local feature vector is extracted by the encoder based on Dilated Convolution Network. The selection gate is in charge of filtering out redundant data. The output of the two encoders is fused as the input of the decoder to improve the feature representation at the source to generate diversified summaries. The model has good performance and effectively enhances the quality of summary, according on the experimental findings on two tough datasets, CNN/DailyMail and DUC 2004.