A Target-Specific Sentiment Classification Method Based on Graph Neural Network

Abstract

The present invention relates to a specific target emotion classification task based on a graph neural network, including collecting data sets and initializing a BERT model, obtaining a one-dimensional feature vector of each target word through the BERT model, and inputting the feature vector of the target word into the graph convolution In the neural network model, the network topology map is constructed, and the adjacency matrix is ​​calculated. According to the adjacency matrix, the three characteristics of the nodes in the network topology map are obtained in three ways, and the relationship classification task is introduced. The whole model is divided into two stages in classification. The two tasks are the sentiment polarity classification of target subjects and the relationship classification between target subjects. The present invention uses a graph neural network to compose multiple subjects appearing in a sentence and process multiple targets at the same time, which is more in line with the cognitive law of human beings to judge emotional polarity, and helps to ensure the effect of the model. At the same time, it introduces The relationship classification task is assisted in classification, which further improves the accuracy of classification.

Description

technical field [0001] The invention relates to the field of natural language processing, in particular to a specific target emotion classification method based on a graph neural network. Background technique [0002] With the emergence of various new forms of Internet media and the development of e-commerce platforms, the growing user groups have generated massive amounts of user-generated content while participating in extensive network activities. In this context, accurate sentiment analysis and opinion mining for massive subjective texts on the Internet becomes particularly important, among which fine-grained sentiment analysis has developed rapidly in recent years. Fine-grained sentiment analysis aims to mine the different emotional tendencies of users towards different target subjects in a comment. In practical application scenarios, there are great application prospects in fields such as refined product models, user portraits, and personalized recommendations. [000...

AI Technical Summary

Problems solved by technology

[0005] The method based on the sentiment dictionary mainly judges the sentiment polarity of the text by constructing a sentiment dictionary and a series of rules, relying on some existing sentiment dictionaries and domain dictionaries and a series of rules to build a sentiment classifier, but the maintenance of the sentiment dictionary requires It consumes a lot of manpower and material resources, and with the continuous emergence of new words, it can no longer meet the application requirements, and needs to be improved and optimized urgently

Or domain experts select a set of statistically significant features from the text, and then use machine learning methods to build a classification model to identify the emotional polarity of the text. Common classification models include naive Bayesian, maximum entropy, and support vector machines, etc. , but the shortcomings are: for different data sets, experts need to select different features, and the investment cost is relatively high. Different feature selection methods will make the final classification results vary greatly, and there may also be differences in different data sets. Large variance in performance, poor generalization of the model

The disadvantage is that the word vectors trained by the Word2Vec and Glove models are static word vectors, which fail to take into account the different meanings expressed by the same word in different contexts and contexts.

At present, the BERT model is rarely used in specific target emotion classification tasks, and in the existing methods, when there are multiple target subjects in a sentence, when the existing method is used to perform emotional classification on a specific target, the model processes the multi-target The emotion classification task is split into single-target emotion classification tasks for processing, ignoring the inherent correlation and connection between different target subjects in the same sentence

Images