Learning How to Self-Learn: Enhancing Self-Training Using Neural Reinforcement Learning

Self-training is a useful strategy for semi-supervised learning, leveraging raw texts for enhancing model performances. Traditional self-training methods depend on heuristics such as model confidence for instance selection, the manual adjustment of which can be expensive. In addition, characteristics of extra training sentences are not considered beyond the baseline method. To address these challenges, we propose a deep reinforcement learning method to learn self-training strategy automatically. Based on the neural representation of sentences and the hidden features from classifiers, a deep Q-network based model is designed to capture their linguistic characteristics and learn an optimal policy for instance selection automatically. Results show that our approach outperforms baseline self-training in terms of better performances and stability.