Bertmucb: An Efficient Approach for Sentiment Analysis

Recent research has made huge progress in sentiment analysis, however, challenges remain in balancing model complexity with performance, also how to train models to better fit into conversations is another problem. This research proposes a novel algorithm for sentiment analysis called BertmUCB. It leverages Reinforcement learning's capability to exploit historical feedback and language models' efficiency to gain better performance, and also considers the feasibility of deploying into reality. The author first fine-tunes a RoBERTa model with a two-layer Multi-Layer Perceptron (MLP) classifier on top of the simplyweibo_4_moods dataset, then, a prediction enhancement method optimized by a Proportion-Integration-Differentiation (PID) controller is applied to the MLP's output. PID adapts parameters in prediction enhancement to better fit different datasets. After prediction enhancement, the author uses a modified Upper Confidence Bound (UCB) algorithm to perform the arm selection. In the experiment, BertmUCB outperforms language models, especially on datasets where language models struggle to distinguish fine-grained sentiments, and an ablation study is conducted to quantify the contribution of each module.