摘要 : The sentence-level extracted summary is inevitably mixed with redundant information due to the uninformative phrases or detailed expressions mixed in it. The extraction of fine-grained units is dedicated to retain the semantical i... 展开 The sentence-level extracted summary is inevitably mixed with redundant information due to the uninformative phrases or detailed expressions mixed in it. The extraction of fine-grained units is dedicated to retain the semantical integrity. To keep the balance between text redundancy and semantical integrity, we propose a novel clause-aware summarization model (TDCL-ClauseSum). We separate complex sentences into grammatically independent but semantically dependent clauses. The clause is regarded as the extraction unit and leverage graph neural network and topical information to capture clause-level relationship. Then a decoupled contrastive loss is stacked over the neural model to fill the gap between topic prediction and clause classification. The experiments of TDCL-ClauseSum are evaluated on two public benchmark datasets CNN/daily mail and New York Times, which contain 310574 and 150536 samples, respectively. Various experiments show that our method achieves remarkable performance on the two datasets (CNN/daily mail:43.94/20.65/40.75, New York Times:49.69/29.84/43.01, in ROUGE-1/ROUGE-2/ROUGE-L). Its promising performance demonstrates that the superiority of clause extraction. 收起