基于深度学习的自动文摘句排序方法

Deep learning approach for automatic abstract sentence ordering

针对现有的文摘句排序方法难以理解深层语义的问题,提出一种基于深度学习的多文档文摘句排序方法。设计端到端深度神经网络完成语句的嵌入、理解及排序。用循环神经网络对句子进行单词级嵌入,在此基础上构建句子的上下文向量表示,用RNN对句子在不同位置的内聚性进行评估,利用指针网络RNN进行下一句预测。实验结果表明,相比传统方法,采用该方法能够得到更高质量的多文档文摘,在自动文摘生成及自然语言处理等方面有广泛用途。

Automated ordering of sentences from multiple documents is a major open challenge in NLP research.To address this problem,an end-to-end deep learning based approach for sentence ordering was proposed.Sequence-to-sequence framework was used,and sentences were embedded into vectors.With the vectors,the assignment of sentences at k positions was predicted using a scoring function.Experimental results show that the proposed model outperforms state-of-the-art approaches in sentence ordering tasks.Results indicate wide applications on sentence ordering,abstract generation,and natural language understanding.