A Data Distribution-driven Sentiment Classification Method Based on Few-shot Gradient-less Learning
This paper describes a few-shot gradient-less learning method driven by data distribution,which designs feature channels that can stably extract effective information during the continuous mapping of data distribution space through optimal transmission theory and Granger causality,and applies the capital asset pricing theory to generate a global high-dimensional optimal revunue structure representations,finally generates a sentiment classification model by using Multi-head Self-attention mechanism.We apply the method to the sentiment classification problem,and can improve the accuracy of the BERT model by 7.07％on the general English dataset and 2.23％on the Chinese dataset in this domain.Applying the method to LCF-ATEPC,an emotion classification model that achieves SOTA performance,the accuracy rate has increased by about 0.6％on average.
computer technologyintelligent computinggradientless learningemotion classificationattention mechanismcapital asset pricing model
NETL
NSTL
万方数据
