Lifelong Sentiment Classification Based on Adaptive Parameter Updating

A classifier with the ability to handle continuous streams of opinion information on the Internet should have good lifelong learning ability. However, deep neural networks face the challenge of catastrophic forgetting when continuously incorporating new domain data, resulting in the loss of previously learned information. At the same time, using the knowledge of old tasks to help the learning of new tasks is also a challenge faced by lifelong learning. In this paper, we propose a novel lifelong sentiment classification method based on adaptive parameter update, which effectively prevent the catastrophic forgetting and promote knowledge transfer among tasks. In our method, we use the uncertainty regularization parameter update strategy to prevent the forgetting of old domain information, and propose an effective parameter updating strategy for sub-network, which can be used to realize the knowledge transfer among tasks. Extensive experiments on 16 popular review corpora demonstrate that the proposed method significantly outperforms the strong baselines for lifelong sentiment classification.