基于关键词主题控制的文本生成技术研究

Research on Text Generation Technology Based on Keyword Topic Control

深度学习技术的发展使得基于深度神经网络的方法成为自然语言处理(Natural Language Processing,NLP)领域的一种新解决思路。虽然神经网络技术能够有效提升生成文本的质量,但是生成文本的内容很容易偏离作者原本要表达的语义。笔者按照人们的写作习惯,即先构思出各部分的主题再进行写作的方式,提出基于交叉项编码的关键词主题控制文本生成模型。与其他模型相比,该模型生成的句子不仅在双语互译质量评估(Bilingual Evaluation Understudy,BLEU)中的得分更高,Correlation值也要高于其他模型。

With the development of deep learning technology, the method based on deep neural network has become a new solution in the field of NLP. Although neural network technology greatly improves the quality of the generated text, the content of the generated text is easy to deviate from the theme, that is, the semantics that the author originally wanted to express. According to people’s writing habits, that is, to conceive the theme of each part before writing, this paper proposes a keyword theme controlled text generation model based on cross-item coding. Compared with other similar models, the sentences generated by this model not only score higher in the BLEU index, but also the results of Correlation are better than other models.