融合词汇特征的生成式摘要模型

生成过程中利用词汇特征(包含n-gram和词性信息)识别更多重点词汇内容,进一步提高摘要生成质量,提出了一种基于sequence-to-sequence(Seq2Seq)结构和attention机制的、融合了词汇特征的生成式摘要算法。算法的输入层将词性向量与词向量合并后作为编码器层的输入,编码器层由双向LSTM组成,上下文向量由编码器的输出和卷积神经网络提取的词汇特征向量构成。模型中的卷积神经网络层控制词汇信息,双向LSTM控制句子信息,解码器层使用单向LSTM为上下文向量解码并生成摘要。实验结果显示,在公开数据集和自采数据集上,融合词汇特征的摘要生成模型性能优于对比模型,在公开数据集上的ROUGE-1,ROUGE-2,ROUGE-L分数分别提升了0.024,0.033,0.030。因此,摘要的生成不仅与文章的语义、主题等特征相关,也与词汇特征相关,所提出的模型在融合关键信息的生成式摘要研究中具有一定的参考价值。

Abstractive summarization model considering hybrid lexical features

In order to use lexical features (including n-gram and part of speech information) to identify more key vocabulary content in the summarization generation process to further improve the quality of the summarization, an algorithm based on sequence-to-sequence (Seq2Seq) structure and attention mechanism and combining lexical features is proposed. The input layer of the algorithm combines the part of speech vector with the word vector, which is the input of the encoder layer. The encoder layer is composed of bi-directional LSTM, and the context vector is composed of the output of the encoder and the lexical feature vector extracted from the convolution neural network. The convolutional neural network layer in the model controls the lexical information, the bi-directional LSTM controls the sentence information, and the decoder layer uses unidirectional LSTM to decode the context vector and generates the summarization. The experiments on public dataset and the self-collected dataset show that the performance of the summarization generation model considering lexical feature is better than that of the contrast model. The ROUGE-1, ROUGE-2 and ROUGE-L scores on the public dataset are improved by 0.024, 0.033 and 0.030, respectively. Therefore, the generation of summarization is not only related to the semantics and themes of the article, but also to the lexical features.The proposed model provides a certain reference value in the research of generating summarization of integrating key infromation.