Aspect-level emotion classification model based on multi-memory attention network

Abstract

The invention discloses an aspect-level sentiment classification model based on a multi-memory attention network. The aspect-level sentiment classification model comprises a word embedding layer, a position memory layer, a BiLSTM network layer, an attention interaction memory layer and a label memory layer, characteristic information is learned through a plurality of memory modules, the emotion characteristic vector of the context is generated more accurately, and the model performance is improved.

Description

technical field [0001] The invention relates to the technical field of emotion classification, in particular to an aspect-level emotion classification model based on a multi-memory attention network. Background technique [0002] Sentiment analysis is an important task in natural language processing. It refers to the use of computers and other auxiliary means to judge people's emotions, opinions, etc. feel. In the research field of sentiment analysis, people have been focusing on the problem of aspect-level sentiment analysis, including two sub-tasks of aspect-level sentiment classification and aspect word extraction. [0003] The present invention mainly studies the emotion classification task of the aspect level, that is, for the specific aspect A of the object O expressed by the text sentence S, judge the emotional polarity about A expressed in the text. For example, given the context: a bunch of friendly staff, the pizza is good, but the beef cuts are not worth the mon...

AI Technical Summary

Problems solved by technology

[0006] The above research began to focus on the importance of aspect words for aspect-level sentiment analysis, but there are still some problems in the above model: 1) Lack of deep mining of information in the context of the target aspect words

2) The IAN-based model needs to use the pooling operation to supervise the generation of attention, which will ignore some internal information between the target word and the context

3) The above models often only focus on the emotional characteristics of aspect words from one angle, and cannot consider them from multiple angles

3) Label memory module, we use the label smoothing regularization method to encourage the model to be less confident about fuzzy labels, because the label unreliability problem is easy to be ignored in previous studies, neutral emotion is a kind of vague emotional state , will bring difficulty to model learning

Images