Research On Aspect Based Sentiment Analysis And Methods Of Recommendation System

With the rapid development of e-commerce,the recommendation system plays a more and more important role in online shopping.We have more and more require-ments on the recommendation systems.Recommendation systems are not only required to meet the personalized needs of users,but also to be finer-grained and more inter-pretable.Benefited from the rapid development of e-commerce platforms,the abundant comment information on them can help us realize these requirements,which promotes the development of aspect based interpretable recommendation systems.Current aspect based recommendation systems mainly consist of two major mod-ules.Firstly,the recommendation systems use the technology of Aspect Based Senti-ment Analysis(ABSA)to extract aspect-level information from reviews or generates text representation in different aspect views.Secondly,the aspect-level features and text representation features are added into user preferences and goods proper-ties respectively to enrich representation of users and goods so that recommendation systems can make more accurate and personalized recommendations.There are many problems existing in the current aspect-level recommendation sys-tems in the two modules,including ignoring the problem characters in aspect identifica-tion,the lack of structural information in aspect-level sentiment polarity classification,the fusion of aspect information and recommendation systems,the rough or even lack of interpretability of recommendation and so on.So this thesis proposes the following solutions:1.For the problem of most works paying all attention to models and feature engi-neering for aspect identification,this thesis makes two observations after analy-sis.Firstly,reviews are short and different segments are independent and usually express different aspects.Secondly,some words in reviews have strong indica-tion for aspects.According to the two observation,this thesis proposes a reviews-segmentation-based method to divide a review into multiple segments with depen-dency parsing tree and identify aspects respectively.Also alignment features are added into models to improve the aspect identification performance.2.For the problem of the lack of structural information in aspect-level sentiment po-larity classification,this thesis introduces a text representation method based on reinforcement learning.Policy network is added to decide word deletion opera-tions according to different aspect so that different text representations in different aspect views can be obtained and used to make more accurate sentiment polarity classification.3.For the problem of the fusion of aspect information and recommendation systems and the rough or even lack of interpretability of recommendation,this thesis trains classification models on external standard annotation data and labels reviews in recommendation.Then,abstract representation features of review are obtained by CN-N and aspect identification is used as an auxiliary task to supervise the process of text modeling.These two kinds of features are both added to the recommendation system to enrich user and good representations.At the same time,aspect-level recommendation reasons can be obtained according to the aspect and sentiment labels of users and goods to meet the need for more granular interpretability.This thesis proposes a new aspect identification model based on segments,a new aspect-level sentiment polarity classification model based on reinforcement learning and a new aspect based interpretable recommendation system and experimental results show the effectiveness of the proposed models.Many detailed comparative experi-ments are also performed to prove the effects of proposed modules and algorithms.