Research On Fine-grained Opinion Mining For Target-oriented Aspect

Opinion mining is one of the core tasks in the field of natural language processing.Among them,fine-grained opinion mining can accurately extract opinions from specified targets,which has very important theoretical significance and commercial application value.With the development of deep learning,fine-grained opinion mining by neural network has become a mainstream method.However,the refinement of analysis granularity leads to a sharp increase in the difficulty of neural networks to capture key information,which is especially prominent in the analysis of comment text with multiple different emotional aspects.All kinds of irrelevant information in such complex text cause serious interference to the extraction of opinions of target aspects.In addition,most finegrained opinion mining tasks only carry out effective polarity classification for the target aspects,which cannot fully represent the specific emotional information of the target aspects,resulting in incomplete and incomplete opinion mining.In response to the above problems,this paper starts with how to conduct comprehensive and accurate opinion mining for target aspects,and proposes a finegrained opinion classification method and an opinion extraction method.First,the key information position is indicated by the aspect and text,so as to attach strong attention to the relevant text and realize the purified representation of text.Second,through the fusion of the corresponding emotional category characteristics of the target to enhance the ability of polarity guidance,to achieve the acquisition of explanatory words of opinion orientation,the target aspects constitute a more comprehensive opinion mining.The main contributions of this paper are as follows:(1)This paper proposes a fine-grained opinion classification method which integrates multi-head interactive indication mechanisms.Multi-head interactive attention is calculated between aspect and comment text,important feature information related to both parties is extracted,so as to remove the interference of semantic coding with irrelevant information in text or target.A neutral label smoothing mechanism is designed to add noise to the output to constrain the model,in order to alleviate the untrusted problem of neutral category with implicit opinions and reduce the risk of model overfitting.Bert dynamic word embedding model is introduced to fine-tune text vectors of different contexts to enhance the accuracy of unstructured text conversion.(2)This paper proposes a fine-grained opinion extraction method of joint target emotion polarity.The polarity guidance mechanism is designed to combine the explicit features of the opinion category into the aspect hidden layer,and the information of the polarity category is used to assist the downstream semantic coding.A special bidirectional GRU mechanism and self-attention mechanism are introduced to integrate target and text,in order to extract more effective semantic representations of different target aspects and the same comment text fully.Through the extraction of emotion category reason word,strengthened the opinion mining comprehensiveness.This paper is a comprehensive fine-grained opinion mining research on target aspects.Experimental results show that this method can effectively improve the accuracy of opinion classification and extract opinion words accurately,which constitute a complete information mining.