Research On Aspect-based Sentiment Classification Based On Deep Learning

Aspect-based sentiment analysis(ABSA)is a basic task in the field of sentiment classification.Its main task is to perform sentiment classification on phrases or texts given a target entity.Early completion of this task generally uses an emotional dictionary-based approach and shallow machine learning.In recent years,with the rapid development of deep learning,most researchers have focused more on adopting deep learning methods.The method based on deep learning can be roughly divided into two steps: First,convert the phrase or text to be classified into a word vector.Second,use neural networks to classify the transformed information.In the first step,since the training of word vectors requires a huge data set,previous methods based on deep learning did not spend a lot of time on training word vectors,but directly used the results of pre-training with deep learning frameworks,but these words The vector set cannot cover all the words,and the words in different contexts will have different contexts,which results in the pre-trained word vector is not robust.In the second step,how to more accurately determine the relationship between target entities and sentiment words has always been a problem to be optimized for specific target sentiment classification.In addition,the contextual semantics of long texts are difficult to get accurate preservation problems still exist.In order to improve the robustness of the pre-training and the accuracy of the classification,the main research work in this paper is as follows:First,Aiming at the problem of robustness of pre-trained word vectors,a method of finetuning pre-training is proposed,and the fine-tuned intermediate information is used as the original information for subsequent neural network classification.To improve the robustness of pre-trained word vectors on different task data.Second,It is proposed to divide the target entity and the context into three parts: the above,the target entity,and the following,in order to reduce the length of the context when using the attention mechanism to query,so as to optimize the long text.It is difficult to save the accurate context semantic problem.Afterwards,the three parts are used for cross-attention query,so that the attention is optimized with concentration,so that the model can better determine the relationship between the target entity and the emotional word,and the classification accuracy is effectively improved.Based on the above two points of work,this paper proposes an WV-LRC(left-right-center with word vector fine-tuning)specific target emotion classification model.In addition,in this paper,the WV-LRC model was carried out on Sem Eval2014's two datasets Laptop,Restaurant.And compared with other advanced models,the experimental results show that the model proposed in this paper has indeed effectively improved the accuracy of the specific target sentiment classification problem,and this model was applied to the engineering needs of enterprise public opinion monitoring.At the same time,the corresponding APP development for financial banks was realized.This APP not only reduces the financial bank's investment in Internet information mining,but also greatly improves the speed of Internet information mining.Currently,it is online in major Android application stores.