Research On Aspect-level Sentiment Analysis Algorithm Based On Deep Learning

Sentiment analysis is a very important research direction in the field of natural language processing.Previously more research was on judging the overall sentiment of an article or a sentence is positive or negative,while aspect-level sentiment analysis aims to identify specific objects its emotional polarity in their background,such as sentences: ”I hated their fajitas,but their salads were great” ? {fajitas: negative,salads: positive }.This more fine-grained sentiment analysis will meet the needs of the now popular ”Internet +” economy.For businesses and consumers,they can make full use of feedback from other consumers and learn more about each attribute of the product.Specific sentiment classification to make behavior decisions more clearly.Deep learning has now become the main method for processing sentiment analysis tasks,usually using RNNs to process in a sequence of texts,learning the semantic information of each word to get the final semantic representation,and then performing sentiment classification.However,these methods have the following shortcomings: 1.During the processing,the difference between the target words and the context words is ignored,and the target words and the context words are treated equally;2.The lack of key context words that have a greater impact on the target word's emotion attention;3.Because RNNs need the output of the previous time step to calculate the hidden state of a word,the formation of sequence dependence lacks the ability of parallel calculation;4.The lack of distinction between the importance of different words in the target words.This thesis designs two models for these problems,and implements an aspect-level user comment sentiment analysis prototype system based on these algorithm models.The main work is as follows:1.In order to distinguish target words from context words and to strengthen the focus on key context words that have a greater impact on target words,an attention model based on LSTM structure(AA-LSTM)is proposed.Improve target words representation and concatenate target words and context words as model input,use two LSTM structures to obtain the semantic information of the context from the two ends of the text to the target words direction,and use the attention mechanism to affect the emotional polarity of the target words pay attention to the semantic information of major words.The model conducted two,three and four classification experiments on the Taobao dataset.The accuracy of the two classification experiments is 95.79%,the accuracy of the three classification experiments is 86.45%,and the accuracy of the four classification experiments is 79.73%.It is nearly 3% higher than the other baseline models in the two classification,more than2% higher than the other models in the three classification,and 1.54% higher than the other baseline models in the four classification.The accuracy and F1 value are better than the benchmark model.2.In order to distinguish target words and context words,strengthen the focus on key context words,improve model parallel computing ability,and distinguish the importance of different words in target words,a double-layer attention model(DA-ABSA)is proposed.The model uses attention mechanism instead of RNNs as a feature extractor,which solves the problem of sequence dependence generated by using recurrent neural networks.Fusion of word vector and word position information as model input;splicing multiple self-attention structures for feature extraction,so that the model learns relevant features of input data in different subspaces;using text attention mechanism to distinguish the importance between words in the target words,to obtain more meaningful characteristics of target words,in order to fine-tune the attention weight of context words.The evaluation and comparison experiments were performed with the evaluation task of Semeval-2014 task4.Compared with the benchmark model,the Accuracy,Precision,Recall and F1 values all performed best.The running time is 793 s,the accuracy rate on the Restaurant data set reaches 82.77 %,the running time is 966 s,the accuracy rate is improved by 1-2% compared with the benchmark model under the same experimental conditions,and the running time is saved by more than 30 %.3.Based on the above two chapters of the algorithm models,a prototype system for aspect-level user comment sentiment analysis is designed to simulate the actual scenario.The application of the algorithm model includes data crawling,data processing,sentiment analysis,and results display.The visual display makes the processing flow of the algorithm model more complete.