Emotion classification method and system, storage medium and equipment

Abstract

The invention relates to an emotion classification method and system, a storage medium and equipment. The method comprises: encoding a context by using a position word vector and multi-head self-attention; encoding the target word by using a bidirectional gating circulation unit and multi-head self-attention; fully extracting semantic information and position information of long and short sentences; meanwhile, interactively splicing the context semantic information and the target word semantic information for low-order fusion; performing position coding on the basis of low-order fusion by using a capsule network; and performing high-order fusion on the information after low-order fusion by using a multi-head interaction attention mechanism, averagely pooling the high-order fusion information, and splicing the averagely pooled high-order fusion information with the averagely pooled target word semantic information and the averagely pooled context semantic information to obtain target feature representation. Compared with the prior art, the context semantic information, the target word semantic information and the position information are fully fused, and the emotion classification accuracy and efficiency are improved.

Description

technical field [0001] The present invention relates to the field of natural language processing, in particular to an emotion classification method, system, storage medium and equipment. Background technique [0002] In recent years, with the wide application of the Internet and the rapid development of e-commerce, the sentiment classification of online comments has shown a huge application demand in e-commerce, information security and public opinion monitoring. Specific target sentiment classification (Aspect-based Sentiment Analysis, ABSA) is a fine-grained sentiment classification task in the field of text sentiment classification. Its main goal is to judge the sentiment tendency (positive, negative and neutral) of sentences corresponding to different goals. . For example: Thismobile phone is beautiful in appearance but expensive in price. As far as the specific target appearance is concerned, the emotional polarity is positive; but for the specific target price, its em...

AI Technical Summary

Problems solved by technology

Although these methods can better solve specific target emotion classification tasks, they still face three challenges: (1) When encoding the semantic information of sentences, each output state of RNNs depends on the output of the upper state. Problems such as loss of long-distance emotional information words and inability to perform parallel calculations on input data

At the same time, due to the over-dispersion of the weight value distribution in the standard attention mechanism, it is easy to introduce excessive noise, and it is difficult to accurately extract sufficient contextual emotional information.

(2) When encoding the semantic information of a sentence, most methods ignore the importance of the position information of the target word to the syntactic structure of the context. The introduction of the position word vector can only solve the position information of each word at a shallow level, and cannot be used for the entire context. Syntactic results are dynamically updated and reconstructed

(3) When fusing context and target words, most of the combination methods based on simple splicing or linear multiplication may lose part of the original information, and cannot fully integrate the two information; and only consider the specific target for different components of the context sentence , ignoring the influence of the context sentence on the target entity

Images