An Aspect Level Sentiment Analysis Based On Deep Neural Network

Aspect level sentiment analysis is a more fine-grained task in sentiment analysis.It is an aspect-based sentiment analysis task,and can be applied to the commodity evaluation system of the e-commerce platform.In a sentence,there can be multiple entities,that is,multiple aspects.And the sentiments contained in these aspects may be different.The Aspect level sentiment analysis task is to accurately identify the different sentimental tendencies contained in different aspects in a sentence.This paper has studied the above problems and proposed two algorithms:First,a sentiment analysis algorithm based on embedding fine-tuning(EF-GRU).In the sentiment analysis task,good embeddings should make the similarity between words of same polarity closer;and the similarity between words of different polarity further.Pre-trained word vectors do not fully comply with this requirement.Therefore,the EF-GRU algorithm uses fine-tuning techniques to fine-tune the joint word vector of context words and target words,making the word vector more suitable for sentiment analysis tasks.Second,a global and local attention network on multi-layer GRU(ATTGL-M).The ATT-GL-M algorithm is the improvement of the EF-GRU algorithm.The ATT-GL-M algorithm is based on the attention mechanism.In the global attention calculation,the ATT-GL-M algorithm calculates the attention value between each context word and the target word,and the word with high attention value should be paid more attention.In the calculation of local attention,the algorithm only pays attention to the attention value of several words around the target word,reducing noise and making the calculation more accurate.Further,this paper extends the model to multiple layers to extract low-dimensional and high-dimensional features simultaneously.Both of the algorithms presented in this paper achieved an accuracy of 10% to 25% higher than baseline on the four data sets.The results of the ATT-GL-M algorithm are also 2% to 3% higher than experimental results on other papers.At the same time,the fine-tuned embedding also achieves good results.After fine-tuning,the similarity between words with opposite polarities decreases by 300% to 400%.