Abstract: RNN-based sequence-to-sequence models have achieved good performance on abstractive summarization. However, these models have some shortcomings including repetitive and exposure bias. A model is presented based on mixed attention including temporal attention and decoding self-attention saving history attention and adding attention for the decoded word to optimize repetition problem. Reinforcement learning is used as a new training method to solve the problem of exposure bias, and modifying the loss function to improve the result. The proposed method is tested using CNN/Daily Mail data set by ROUGE, showing that mixed attention can improve the repetition problem, and the exposure bias can be eliminated by reinforcement learning, and the integrated model surpasses the advanced algorithm on the test set.

Key words: abstractive summarization, mixture attention, reinforcement learning, natural language processing, exposure bias, recursive neural network

摘要： 基于递归神经网络的序列到序列的模型在文本摘要生成任务中取得了非常好的效果，但这类模型大多存在生成文本重复、曝光偏差等问题。针对重复问题，提出一种由存储注意力和解码自注意力构成的混合注意力，通过存储历史注意力和增加对历史生成单词的注意力来克服该问题；使用强化学习作为一种新的训练方式来解决曝光偏差问题，同时修正损失函数。在CNN/Daily Mail数据集对模型进行测试，以ROUGE为评价指标，结果证明了混合注意力对重复问题有较大的改善，借助强化学习可以消除曝光偏差，整合后的模型在测试集上超越先进算法。

关键词: 文本摘要生成, 混合注意力, 强化学习, 自然语言处理, 曝光偏差, 递归神经网络