基于复制和覆盖率机制的生成式文本摘要方法研究

Research on Abstract Summarization Method Based on Copy and Coverage Mechanism

生成技术将给定的冗长文本数据中的重点信息提取出来组成文本摘要,有利于信息的搜集和后续处理。以序列到序列模型为基础,模型使用双向长短期记忆神经网络作为编码器单元获取上下文相关隐藏状态向量,针对文本摘要生成领域的长期依赖问题、未登录词问题和目标摘要中字词重复问题,分别使用注意力机制、复制机制和覆盖率机制,在中文新闻短摘要文本数据集LCSTS上进行实验,以ROUGE系列评分作为评价标准,平均提高0.3个百分点。

Automatic text summarization method extracts the key information from a given lengthy text to form a text summary,which is conducive to the collection and processing of information.Based on the sequence to sequence model,the model uses the bidirectional long-term and short-term neural network as the encoder to obtain the context sensitive hidden state vector.For the long-term dependence in the field of text summary generation,the problem of Out-Of-Vocabulary words and the problem of word repetition in the target summary,the attention,copy and coverage mechanisms are used respectively,The experiment is carried out on the Chinese News short summary text dataset named LCSTS,and the ROUGE series scores are used as the evaluation standard,with an average increase of 0.3 percentage points.