| Automatic text summarization aims to obtain a summary which refines information of source text by using the superiority of machines,which is of great significance in solving the problem of information overload in the era of Internet information explosion,and at the same time improving people's work efficiency.Benefiting from the development of the Seq2 Seq model,abstractive summarization has achieved remarkable results and has become a new research hotspot,but there are still two challenges: First,due to the limitation of the length,the generated summary should cover more salient information of the source text in a shorter space.Second,the generated summary should be more abstractive,that is,contain more novel words instead of just copying the words of the source text.In this thesis,we make an in-depth study on abstractive summarization based on the Seq2 Seq model,and propose a Topic-aware Pointer Model(TPM)innovatively to address the challenges in abstractive summarization.Specifically,the topic information mined from source text is incorporated into TPM as prior knowledge from two aspects: attention mechanism and pointer mechanism,which helps the TPM better to attend the salient information of source text.Besides that,the addition of topic information allows the TPM to identify novel topic words that do not appear in the source text,which increases the abstraction of the generated summary.Further,we introduce a novel Topic Relevance Loss(TRL)function during the training stage to encourage the topic similarity between source text and generated summary.Finally,we conduct comparative experiments on the real scientific datasets in this thesis.The experimental results show that TPM with TRL function has achieved the best results in the ROUGE evaluation standard compared to the existing abstractive summarization model,and the generated summary with more abstraction and higher topic similarity,also covers more salient information of source text. |
PDF Full Text Request