Study On Text Summarization Based On Neural Network Model With Selective Mechanism

Nowadays,people can receive a great deal of information from multiple sources all the time,and how to figure out the important information becomes particularly important,which can be achieved by text summarization.We mainly focus on the sentence-level abstractive text summarization task and the document-level extractive text summarization task.For the task of sentence-level abstractive text summarization,two Sequence-to-sequence(Seq2Seq)models are proposed: one is based on Recurrent Neural Networks(RNN),and the other is based on Convolutional Neural Networks(CNN).For the RNN-based model,a joint selective mechanism is proposed to select the salient information by reading the input and the output jointly.For the CNN-based model,a global encoding mechanism is introduced to generate a new representation of each input word with the global context.For the task of document-level extractive text summarization,an unsupervised model and a supervised model are proposed.The unsupervised model is based on the TextRank model and is calculated according to the sentence semantic similarity.The supervised model is based on the Seq2 Seq framework,and a procedure similar to the Pointer Network is adopted.A selective mechanism is also introduced to select important sentences more accurately?In experiments,for the task of sentence-level abstractive text summarization,the proposed model outperforms the Seq2 Seq model with 1.92 ROUGE-2 on the commonly used English Gigaword dataset.For the task of document-level extractive text summarization,the CNN/Daily Mail dataset is used as the test set.The proposed unsupervised model outperforms the TextRank model with 0.94 ROUGE-2.The proposed supervised model outperforms the baseline model with 0.85 ROUGE-2.We also implement a text summarization system based on the proposed models.