Research On Conditional Text Generation Technology And Its Application Based On Deep Learning

In recent years,with the rapid development of network information technology,the spread of network information has subverted the traditional social public opinion environment and gradually possessed the leading position.However,many negative and socially harmful news has widely spread relying on the convenience of the Internet.Therefore,in order to build a healthy online public opinion environment,it is urgent to strengthen the guidance of online public opinion.One of the important issues is the automatic generation of online comments.However,the traditional rule-based generation methods are not intelligent and effective enough.So,this thesis studies conditional text generation technology and its application based on deep learning.The main research results are listed as follows:(1)Sentiment controllable conditional text generation is studied.Aiming at handling the attributes entanglement problem that occurs in current text generation methods based on sentiment transfer,an emotional controllable conditional text generation based on sentiment transfer is proposed to alleviate the disfluency of text generation.This method generates a new text with a target emotional value by transferring the emotional value of a given text.Firstly,a large-scale sentiment dictionary is used to match sentiment words in sentences.Secondly,the matched words are replaced with "mask" symbols.Finally,the MaskAE is used to generate the replaced sentiment words,and the other words are kept unchanged,thereby alleviating the problem of text attribute entanglement.The experimental results show that this method can keep the content of the text unchanged,and obtain a higher sentiment controlling rate for the text generation,and effectively improves the fluency of the generated sentence.(2)Sentiment and topic controllable conditional text generation is researched.Aiming at the problem that current methods expect to use deep learning models to infer sentence-level sentiment-independent content representation,a conditional text generation method based on topic keywords is put forward.Firstly,three topic keyword extraction schemes based on part-of-speech categories are designed to extract topic keywords that do not contain emotional information in the sentence.Secondly,the embeddings of topic keywords are used as input to the generation model so that the model no longer needs to infer sentiment-independent content representations,thereby avoiding an unstable adversarial training process.Thirdly,the text generation model is trained with a single-layer LSTM based Auto-Encoder model.Finally,according to the given topic keywords and target emotional value,sentiment and topic controllable text is generated by the trained model.The experimental results show that this method breaks through the traditional sentence-to-sentence generation style by introducing topic keywords,can freely customize the topic and emotion of the generated text,and is suitable for long text conditional generation.(3)The generation of online comments incorporating topics,sentiments and netizen cultural style is explored.Aiming at the special need of online comments for public opinion guidance,that is,topics,sentiments and netizen cultural style should be integrated into online comments.On the basis of researches above,an online comment generation method based on deep learning is presented.Firstly,training corpus is constructed by using online comments related to specific public opinion events to obtain event topic information and netizen cultural style information.Secondly,according to whether the positive comments and negative ones in the corpus are balanced,conditional text generation method based on topic keywords and random generation method based on variational Auto-Encoder are used to train the online comment generation model,respectively.Finally,online comments with topics,sentiments,and netizen cultural style incorporated,are generated by the trained model.The experimental results show that this method can effectively improve the effect of online comment generation and meet its engineering requirements.