Research On Automatic Generation Of News Comment Based On Deep Learning

The automatic generation of news comments is a challenging and under-researched task in natural language processing.With the explosive growth of information on the Internet,people's demand for automatic news comment generation systems has become stronger.It helps to increase user engagement and interactivity on online news platforms,and at the same time can act as a comment assistant and enrich the feature list of chatbots.Last few years,with the rapid development of deep learning,the automatic generation of news reviews has transitioned from a retrieval model to a sequence-to-sequence model based on a deep neural network.However,because news comment automatic generation is different from other text generation tasks,it involves a variety of cognitive abilities,such as understanding articles,forming opinions and arguments,and organizing expression language.Therefore,there are still some research problems to be decided urgently in existing research work.For these problems,this thesis searchs the automatic generation method of news comments based on deep learning,and conducts research on the existing sequence-to-sequence models and algorithms to improve the performance of the automatic news comment generation system.First,this thesis proposes a hybrid encoder-keyword mechanism model for the two problems that the current work does not distinguish between news headlines and news texts,and the existing models cannot fully extract the key information of the news.Specifically,for the first problem,this thesis proposes a hybrid encoder based on bidirectional LSTM modeling encoding for news title and news text,respectively,in order to alleviate the asymmetrical length and information content of the two,so as to improve the quality of text vector representation.In response to the second question,referring to the implementation of the pointer generation network,this thesis designs a keyword mechanism,which includes two parts: keyword attention mechanism and keyword copy mechanism.Through this mechanism,the model can be integrated into the key information in the news to generate keyword-based guided comment.Finally,this article combines the two to form a hybrid encoder-keyword mechanism model.Based on multi-dimensional experimental analysis results show that our model can better obtain the semantic representation of news,improve the performance of the news comment automatic generation system,and the generated comments have a good correlation with news.Second,this thesis aims at the problem that the news text is too long and has complex context information,and the existing work does not model the content structure of the news well,and proposes a topic encoder model.Specifically,first use the LDA topic model to extract and construct the topic of the news text,and convert the unstructured news text into topic-sentence pairs.Then referring to the implementation of the Transformer model,a topic encoder is designed.The topic-sentence pair is input into the topic encoder,and the sequence embedding module,the sentence encoding module and the topic encoding module are used to calculate the hidden state sequence of the topic.Finally,this thesis designs a topic attention mechanism,so that different attention levels are given to each topic in the decoding part to generate diverse comments.Based on multi-dimensional experimental analysis results show that our model can analyze and understand the relationship between content and topic well,generate comments related to news topics and have information content and diversity,and improve the performance of the automatic news comment generation system.Thirdly,This thesis proposes an emotional controllable model to solve the problem of unpredictable sentiment of comments output by existing models.Specifically,firstly,emotional features are extracted by emotional classification of comments,this thesis divides comment sentiment into three categories: positive,neutral and negative.Then the emotional features are integrated into the decoder to guide the generation of comments.Experimental results indicate that the model can achieve the goal of emotional freedom and controllability.