Fine Grained Sentiment Analysis Of Clothing Product Comments Based On Deep Learning

In the Internet age,E-commerce is booming,and the derived product reviews bring in different effects to the businesses and the consumers.For the businesses,they improve and design the follow-up products based on the feedback of the consumers;and for the consumers,they will find the satisfied products referred to the feedback of the other consumers.Text sentiment analysis is one of the commonly used means to study the product reviews.According to the different levels of the analysis objects,it can be divided to the coarse-grained emotional analysis and the fine-grained emotional analysis.Compared with the coarse-grained commodity comment sentiment analysis,the fine-grained commodity comment sentiment analysis can understand the advantages and disadvantages of the goods more specifically,provide a strong reference for the buyer,and facilitate the seller to improve the product quality and shape,which is one of the key and difficult points of the text analysis.The key of the sentiment analysis of the fine-grained product reviews lies on the accurate extraction of the opinion targets and their related descriptions.At present,there are some problems such as the low extraction rate of the low-frequency opinion targets and the difficulty in finding the local sentiment words.This paper proposes a method of extracting the opinion targets and the sentiment words in the clothing reviews combined with the external knowledge to solve the above problems.The difficulty of the sentiment analysis of the fine-grained product reviews lies on the accurate sentiment classification of the related description of the opinion targets.The effect of the in-depth learning method is better,but it is difficult to learn the further features of the text.This paper proposes a fine-grained sentiment analysis model based on the integrated multi-layered feature based on the long short-term memory network to solve the above problems.(1)In this paper,we propose an association combination extraction method of the sentiment words in the clothing product comments based on the external knowledge: it brings in the conceptual hierarchy tree to detail the clothing design to the middle opinion targets,such as modeling,material,color and workmanship etc.and 207 kinds of the bottom opinion targets,like body-fit,sexy,elasticity,texture,process etc.It is to use the Chinese Synonyms to expand the middle and the bottom opinion targets,which has established four middle opinion targets dictionary.A same sentiment word may have several emotional orientations because of the different age,different season and different buyers.Therefore,it is to introduce the multi-oriented sentiment dictionary into the clothing sentiment dictionary.It is to break the product reviews into several short sentences in accordance with the punctuation,and then to use the part-of-speech rules,dependency syntax and feature dictionary to extract the opinion targets and the sentimental words,and then to use the classic association combination to engrave their relationship.It is to have an experiment on the clothing reviews data set and the result shows that the overall accuracy rate of the extraction to the opinion targets and the sentimental words can be up to 89.2%.(2)This paper gives the MLFF-Bi LSTM model(Bi LSTM Based on Multi-level Features Fusion),which is the long short-term memory fine-grained sentiment analysis model integrated with the multi-layered feature.It is to learn the clothing reviews' further feature through the in-depth learning method.And the further feature is formed by the emotional orientation feature and the contextual feature.(1)It is thought that the emotional orientation is not only related with the emotional words but also the opinion targets,hence,it is to extract from both of them.The emotional orientation feature extraction: It is to send the sentence vectors of the reviews into the bidirectional long short-term memory to learn features,and then to gather the word vectors to the opinion targets from the association combination to obtain the attention weight so that to achieve the feature selection;The emotional orientation feature extraction method based on the emotional words: It is to join the negative words,adverbs of degree and emotional word vectors from the association combination to form the input vectors and then to send the input vectors into the bidirectional long short-term memory to learn the features and then to establish the attention mechanism through the max-pooling to select the critical features.(2)The contexture feature extraction: It is to joint the word vectors of the reviews and the combination vectors consist of the opinion target,adverbs of degree,negative words and emotional words and then to send the joints into the bidirectional long short-term memory to study,and then to establish the attention mechanism through the mean-pooling to extract the related overall feature.The MLFF-Bi LSTM model(Bi LSTM Based on Multi-level Features Fusion)is to vectorize the texture of the reviews through the Bert model(Bidirectional Encoder Representation from Transformers),and then to integrate the external knowledge to extract the association combination of the clothing reviews.It is to combine the Bert vector and the association combination and to use the bidirectional long short-term memory and attention mechanism to learn the further feature of the texture,and then to integrate the learned further feature to achieve the sentiment analysis task.It has been made the experiment on the clothing product review data sets established on the texture.The experiment shows that the MLFF-Bi LSTM model can adequately learn the emotional orientation feature and the contexture feature from the reviews,whose effectiveness is better than that of the traditional CNN and LSTM model so that to improve the sentiment classification effectiveness and the model explanatory.