Aspect-Category Based Sentiment Analysis Enhanced by Label Prior Knowledge
Current aspect-category based sentiment analysis aims at performing joint aspect category detection and category-oriented sentiment classification.To deal with the noisy labels often occurred in sentiment datasets,we propose an aspect-category based sentiment analysis approach with label prior knowledge(AP-LPK).Specifically,we firstly construct an autoregressive prompting learning that can effectively utilize the learning ability of pre-trained language models.And then label words are generated through autoregression for better semantic consistency than non-autoregression.Secondly,we introduce the label distribution of each category as label prior knowledge that is refined through Bernoulli distribution to mitigate the interference of noisy labels.And then,the output labels from the autoregressive prompting and label prior knowledge refinement jointly decide the prediction based on the distributions of sentiment polarities.Evaluated on five datasets including the four benchmark datasets from SemEval 2015 and 2016 and the Restaurant-domain dataset from AI Challenger 2018,the proposed approach outperforms existing baselines in terms of F1.
aspect-category based sentiment analysisprompt learninglabel prior knowledge
NETL
NSTL
万方数据
