Research On The Natural Language Processing-based Algorithms For Review Analysis

With the development of electronic information industry,people's life has become increasingly dependent on electronic equipment and Internet.In order to collect the feed-backs from users,a variety of apps have set up comment services.These comments con-tain a lot of valuable information.With this information,businesses can better understand users' hobbies and make recommendations for users.Furthermore,they can better under-stand their own shortcomings and improve their services.However,in the face of mas-sive reviews,businesses or platforms often don't have enough manpower to read them.Therefore,natural language processing technology based on deep learning and machine learning is used to address the issue mentioned above.This thesis focuses on how to use fine-grained sentiment analysis technology to mine comment information and how to use comment information to enable recommendation algorithm.Fine grained sentiment analysis aims to mine the sentiment polarity of specific enti-ties or categories in text information.Traditional sentiment analysis only focuses on the emotional polarity of complete sentences,but in many reviews,different entities have dif-ferent emotional polarities.If we obtain the emotional polarity of each entity,we can pro-vide more accurate and detailed feedback for businesses or platforms,so as to guide them to improve their services more specifically.A common drawback of the existing methods is that the model tends to over associate the label with one side of the comment text or entity category and ignore the other side.Inspired by the control variable method,this thesis proposes a pairwise pretraining method to help the model more evenly distribute its attention on sentences and entities,so that the model can pay attention to the relationship between the two parts of the input data at the same time,and not pay too much attention to one side and ignore the other side.In addition,this thesis also uses PGD adversarial training and extended auxiliary sentences to improve the training effect.Recommendation algorithm can help people find the items they may be interested in.Most of the data used in recommendation algorithm is user's behavior feedback,but the implicit feedback contains fuzzy information,which limits the performance of recom-mendation system.A large number of explicit feedback contained in comments can solve this problem well.Users often explicitly point out what they like or dislike in comments.NLP technology can learn the interests of each user and the attributes of each item through these comments.Therefore,NLP based recommendation algorithm emerges as the times require.Aiming at the data deviation of the existing methods in the training and veri-fication stage,the attention distribution guide information transfer network designed in this thesis uses a teacher student architecture,takes two groups of attention distribution as teacher signals,and transmits the information of the auxiliary module to the main module,which can not only avoid data leakage,but also maximize the use of effective data,and also realize the decoupling of user items at the teacher signal level.The experimental results show that the two algorithm models designed in this thesis are better than existing algorithm models.