Emotion feature representation learning system and method based on global interaction and syntax dependence

Abstract

The invention discloses an emotion feature representation learning system and method based on global interaction and syntactic dependence. The emotion feature representation learning method comprisesthe steps: through the information interaction of contexts and evaluation objects, introduction of a word syntactic dependence relationship and feature joint learning, employing a pre-trained multi-layer language model for carrying out embedded representation on the words, and obtaining more contextualized representation of the words; solving the problem of insufficient interaction caused by independent modeling of contexts and evaluation objects in a previous method by utilizing a superior lifetime double-sentence task mode of a BERT structure. According to the method, a graph dependence attention network is provided, the problem that the graph dependence attention network does not distinguish influences of different dependence relations is solved; meanwhile, the syntactic relation of comments is reasonably modeled into a model, and words are made to represent syntactic dependence information; according to the method, a feature joint learning method is used, interaction information and syntax dependence information of comments are fully combined, and therefore the accuracy of emotion analysis of the evaluation object is improved.

Description

【Technical field】 [0001] The invention belongs to the field of natural language processing technology and emotion judgment, and relates to an emotion feature representation learning system and method based on global interaction and syntax dependence. 【Background technique】 [0002] With the rapid development of the Internet, online shopping has become an indispensable part of people's lives, and the online review data of online products generated by online shopping has also shown an exponential growth. Most of these review data are the real feelings and objective evaluations of consumers after using the product, which can not only guide or promote other consumers’ purchasing interest, but also help product providers to obtain problems, defects and deficiencies in the product, and promote product design and service. Therefore, the mining and utilization of online review data contains important commercial value. Specifically, from the perspective of consumers, what a consumer...

AI Technical Summary

Problems solved by technology

Among them, the template matching method refers to classifying common evaluation objects and the appearance of evaluation words into different templates according to the dependency relationship between words, and using the template to extract candidate evaluation collocations from comments , and determine the emotional polarity of the evaluation object according to the emotional polarity of the evaluation words. This type of method is an unsupervised method. Although it has achieved certain results, it is difficult to deal with more complex sentences and will introduce a lot of noise.

The method based on the emotional dictionary refers to extracting all the words with emotional polarity in the comments through the emotional dictionary, assigning different scores to each emotional word, and finally using the total score of the emotional words in the review as a criterion for distinguishing emotional polarity , this method cannot handle the situation where there are evaluation objects of opposite emotional polarity in the sentence at the same time, resulting in a low accuracy rate

The method based on deep learning mainly uses the ability of automatic feature engineering of deep neural network to model the context of comments and evaluation objects respectively, and uses the interaction of the two to obtain the final emotional features, and uses relative position weights for feature screening. This method is currently the mainstream method of sentiment analysis based on evaluation objects. Although it has made a lot of achievements, there are still attention mechanism noises, information attenuation caused by relative position weights, insufficient interaction between context and evaluation objects, and insufficient word dependencies. Utilization and other issues, which largely affect the accuracy of product evaluation object sentiment analysis

Images