Adaptive deep learning for entity disambiguation via knowledge-based risk analysis

Abstract

The state-of-the-art performance on entity disambiguation has been reached by deep neural networks. However, the task remains very challenging due to the complexity of natural language. Moreover, the target data distribution is often different from that of training data. In this paper, we address the limitation of deep entity disambiguation from the perspective of misprediction risk. We propose a knowledge-based approach of risk analysis for entity disambiguation, and leverage it to enable adaptive deep learning. The proposed approach generates risk features by extracting evidences from the knowledge base, and then models them as a linearly-weighted random vector where an attention mechanism is used to focus on the most significant components. Finally, it estimates misprediction risk of the aggregated probability distribution via the Conditional Value-at-Risk metric. Furthermore, we demonstrate how to utilize risk analysis results in adaptive deep learning via two-phase training, the first phase fits on labeled training data while the second one minimizes misprediction risk on unlabeled target data. We evaluate the performance of the proposed approach on benchmark datasets through a comparative study. Our thorough experiments demonstrate that it can detect mispredictions more accurately than existing alternatives and can substantially improve the performance of deep learning models.