基于IF-PGN模型的短文本摘要生成

Short Text Summarization Based on IF-PGN Model

针对以往提出的模型生成的摘要准确性不够,且含有过多冗余信息的问题,提出一种生成式文本摘要模型--信息过滤-指针生成网络。该模型对神经网络编码解码结构进行扩展,引入信息过滤网络和指针生成网络,可以有效地捕获原文信息,免于无效信息的影响,且让指针网络的复制概率更加准确,生成的摘要内容更加丰富、连贯。在CNN/Ddily Mail数据集上的实验结果表明, ROUGE指标有明显提升。

Aiming the problem of insufficient accuracy and too much redundant information in the abstract generated by previous models, an information filter-pointer generation network is proposed.This model extends the encoding-decoding structure of neural network, introduces information filtering network and pointer generation network, which can effectively capture the original information and avoid the influence of invalid information, and make the replication probability of pointer network more accurate, and the generated abstract content more rich and coherent.The experimental results on the CNN/Ddily Mail dataset show a significant improvement in ROUGE.