Research On Aspect-level Sentiment Analysis Method Based On Multi Task Learning

People often express their personal views,attitudes and sentiment about certain things in social media such as blogs,forums,online reviews,e-commerce platforms,etc.These comments are important resources for quasi consumers to make decisions,and also important basis for product or service providers to draw opinions.The purpose of this paper is to extract the explicit and implicit evaluation objects,opinion words and further judge the sentiment expressed by users.This paper proposes a multi task joint learning method to capture semantic association information among different tasks,which effectively improves the accuracy of system recognition.This kind of fine-grained aspectlevel sentiment analysis research has important academic value and application value for sentiment analysis task.The main research work of this paper is as follows:1.Aspect-level polarity classification method based on improved memory network.Memory network only pays attention to the simple semantics of word level and ignores the semantic information of sequence itself when solving the sentiment classification of evaluation object.In this paper,the convolution multi-head self-attention module is used to extract memory,and then the vector representation of the evaluation object is used to calculate the multi hop attention,in order to make up for the shortcomings of memory network.Our method is to surpass all baseline methods in four datasets(Sem Eval-2014 Restaurant(AT),Sem Eval-2014 Laptop,Sem Eval-2014 Restaurant(AC)and Sem Eval-2016 Tweet),and the accuracy of the four datasets increased by 3.10%,2.91%,1.83%,2.02% respectively.The pre task of sentiment polarity classification for evaluation object is the recognition of evaluation object,including the extraction of aspect terms and the recognition of aspect categories.In this paper,the methods of opinion terms extraction and aspect categories detection are studied by using sequential annotation model and multi-label two classification model of joint learning.2.Aspect based sentiment analysis method based on multi task joint learning model.In most research,the recognition and sentiment analysis of evaluation objects are considered as independent tasks.However,the results of parallel tasks are often needed under the driving of application requirements.Pipeline processing will lead to error accumulation,and can not learn some important cross task correlation information.This paper proposes a joint solution of related sub tasks,which can solve multiple tasks of application requirements in one stop.Specifically,it includes three aspects: simultaneous recognition of aspect categories and aspect terms;simultaneous judgment of aspect categories and their sentiment polarity;simultaneous extraction of aspect terms and their sentiment polarity.The research focus of this paper is to solve different tasks at the same time,and improve the performance of each task by using the correlation information between tasks.The experimental results show that our method is superior to all the baseline systems in the Sem Eval-2014 Laptop dataset,and the F1 value of aspect terms extraction task is 0.54% higher than that of baseline method,and the F1 value of aspect terms and their sentiment polarity judgment task is 0.28% higher.3.The method of aspect-level sentiment analysis which integrates the information of opinion words.In aspect based sentiment analysis,people mainly focus on the evaluation object and sentiment,and opinion terms are also important emotional elements.There may be mutually dependent syntactic relationship between opinion words and aspect words,and opinion words are the important basis for judging sentiment expression.This paper proposes a one-stop solution for aspect word and opinion word extraction and sentiment polarity judgment of aspect word.It tries to improve the performance of each task by using the dependency among three tasks.Compared with the baseline model,the F1 values of three datasets(Sem Eval-2014 Laptop,Sem Eval-2014 Restaurant and Sem Eval-2015 Restaurant)were increased by 3.49%,0.81%,3.26% respectively.In the laptop dataset,the task of aspect word and sentiment polarity judgment increased by 2.42%.