An Empirical Study On Emotion Classification And Joint Entity Relation Extraction Based On Deep Learning

With the rapid development of "Internet + tourism",China's online booking hotel scale is rapidly expanding.Consumers can publish their experience and evaluation on the hotel reservation platform.It has become an important way for users to make reservation decisions and hotel managers to make management decisions,evaluate their own service quality and explore consumer needs.In the hotel review,it includes consumers' emotional tendency to evaluate the service quality of the hotel,as well as an evaluation of the specific evaluation object.Analysis,processing,induction and summary of these texts are of great reference significance to user analysis,market survey and exploration of factors affecting the quality of hotel service.In this paper,a total of 104,740 comments were collected from 88 hotels on ctrip,and emotional classification and relationship extraction were conducted.Emotional classification was conducted on the text,and the evaluation objects,evaluation words and their corresponding relations were extracted.For the construction of emotion classification model,this paper manually annotated the comments and divided them into positive emotion and negative emotion,in which the ratio of positive and negative samples was 9:1.Focal Loss was used instead of cross entropy as the loss function of the emotion classification model,which could reduce the weight of simple samples and focus on the training of difficult samples.This article constructed three different network structure model,Text_CNN model,the hybrid neural network model and ensemble learning FastText model.FastText model is best model in the single model,the F value is 0.8587,by ensemble learning the F value up to 0.9003,AUC value is 0.9295.Compared with single model,ensemble learning model can improve the classification effect to some extent.In this paper,EA+BIOES annotation system is proposed to annotate the evaluation objects and evaluation words in the comment text as input,train them with Bi-LSTM+CRF model,convert extraction task into sequence annotation task withannotation strategy,eliminate the accumulated error of traditional entity identification before relation extraction,and realize JointEntity and relation extraction.The F value of the first evaluation object in the sentence can reach 0.8387,and the value of the overall entity is0.7258.This paper proposes an effective method to extract consumers' viewpoint and entity emotion knowledge from hotel reviews on a large scale,which provides ideas for further constructing hotel evaluation information knowledge spectrum.