Continual Learning with Knowledge Transfer for Sentiment Classification

This paper studies continual learning (CL) for sentiment classification (SC). In this setting, the CL system learns a sequence of SC tasks incrementally in neural network, where each task builds classifier to classify reviews particular product category or domain. Two natural questions are: Can transfer knowledge learned past from previous new help it learn better model task? And, can old models be improved process as well? proposes novel technique called KAN achieve these objectives. markedly improve accuracy both and via forward backward transfer. The effectiveness is demonstrated through extensive experiments (Code data are available at: https://github.com/ZixuanKe/LifelongSentClass).