Research On Aspect Extraction Method In Text Sentiment Analysis

Sentiment analysis is one of the most active research fields in Natural Language Processing.As a fine-grained method,aspect-level sentiment analysis can dig out the tendencies of different aspect from the text and reflect the complex features of emotions.Aspect extraction is the basic work of aspect-level sentiment analysis.At present,in aspect extraction methods aspect words are not closely related to the context,and lack of full use of global information.In addition,few data sets are available and each data set contains a small amount of data.Solving the above problems has a positive effect on improving the effect of aspect extraction.Aiming at the above two problems and incorporating self-attention with the model of Double Embeddings and CNN-based model,this thesis proposes a double embeddings with self-attention and cnn-based method and a text enhancement method based on synonym substitution.The thesis work was supported by the National Key R&D Program Project "Internal and External Connected Trial Execution and Litigation Services Collaborative Support Technology Research"(2018YFC0831300).The main work is as follows:(1)To solve the problem that aspect words are not closely connected with context and lack of global information,double embeddings with self-attention and cnn-based model which is based on the DE-CNN is proposed.By using the self-attention and position encoding,the path length between aspect words and other words is reduced to 1which strengthens the connection between aspect words and other words as well as captures long-distance dependence and more global semantic information.The feature representation of aspect words is enhanced according to the degree of correlation between words,and the eigenvalue of non-aspect words is weakened.This method can provide more useful relevant information for the subsequent learning of Convolutional Neural Network.Experimental results show that the proposed method improves the performance of aspect extraction.Compared with DE-CNN,the average F1 score on the two datasets are increased by 0.52 and 1.74 respectively.(2)To address the problem of few available Data sets for aspect extraction,this thesis proposes a synonym replacement method to augment data which is based on Easy Data Augmentation technology.This method combines with the feature of aspect extraction that label and input should be strictly corresponding,improves the synonym replacement algorithm by adding data preprocessing and expands the train data.Experiments on the augmented data show that the enhanced data can effectively improve the effect of aspect extraction of each model.Compared with the original data,the average F1 score of double embeddings with self-attention and cnn-based model is improved by0.61.The average F1 score of some models is increased by more than 1 compared with the original data.Combined the text augmentation method based on synonym replacement and the double embeddings with self-attention and cnn-based method,experiments show that compared with the DE-CNN model,the average F1 score on the dataset is increased by1.13,which proves the effectiveness and rationality of the two methods.