Attention Networkd For Aspect-Level Sentiment Classification

With modern mobile devices and 5G communication entering thousands of households,traditional offline shopping is gradually penetrated by online shopping.Sentiment analysis of reviews on e-commerce platforms is beneficial to merchants to optimize their products and services,and is also beneficial to buyers to make choices.Different from traditional coarse-grained text analysis,aspect-level sentiment classification aims to identify aspect words and corresponding different sentiments in the text.Therefore,the research of aspect-level sentiment classification mainly focuses on extracting semantic features of the text and establishing the relationship between aspect words and sentiments.First of all,in the problem of aspect word extraction,there is a problem in the definition of aspect words,and the problem of nested aspect words in aspect words is prone to occur,which needs to be solved by using the attention mechanism.Secondly,in the sub-task of sentiment classification,it is often easy to ignore the interactive information between the aspect words and the text,which may misjudge the specific meaning of the text,and then lead to the wrong result of sentiment classification.In view of these existing problems,this paper makes corresponding improvements on the existing work,including the following research contents:1.Before the task of aspect-level sentiment classification,we first need to extract aspect words from reviews.Therefore,this paper proposes an ABSC model for emotional aspect word extraction based on self-attention mechanism.The ABSC model uses ALBERT for word embedding representation,and uses the bidirectional GRU structure for context information feature extraction.Then,the optimal sequence at the vector level is found through the self-attention module.At the same time,the model proves the effectiveness of ALBERT in reducing the training time and the effectiveness of BiGRU in context feature extraction through ablation experiments.Finally,experiments show that the algorithm in this paper improves the F1 value by 3% compared with the baseline model on the relevant data set,which proves the effectiveness of the ABSC model in the aspect word extraction task.2.This paper proposes a BiIAGRU-BERT sentiment classification model with interactive attention to classify the sentiment polarity of a given aspect word and text.Firstly,the model uses BERT to embed words,and then uses the interactive attention mechanism to calculate and capture the interactive information between the context and the aspect words.Secondly,the multi-head selfattention mechanism is used to integrate the different parts of the information,and finally the corresponding sentiment polarity is output.In the experiment of SemEval2014 dataset,the accuracy of the model is improved by 5% compared with other models.3.Based on the model proposed above,an online review sentiment analysis system for ecommerce industry is developed with the help of a common web development framework,and the end-to-end aspect-level sentiment classification task is completed for different aspects of e-commerce user reviews.The system requirements and system architecture are expounded in detail,and the related pages are tested to verify the effectiveness of the system.