Interpretable named entity recognition via integrating logical rule learning with deep neural networks

Deep neural networks have been grappling with limited interpretability, whereas logical rule learning methods offer clear interpretability. Research has shown that neural networks adapting logical rules perform remarkably well and offer interpretability in visual and graph-related tasks. However, in language-related tasks, recent advancements in logical rule learning have mainly concentrated on text classification. Exploring the fusion of logical rule learning and neural networks in tasks like named entity recognition (NER) remains uncharted territory. Hence, to bridge this gap, we propose a new method called knowledge-aware deep logic learning (KDL) explicitly designed for NER, emphasizing interpretability. We delve into implementing KDL for bidirectional recurrent neural networks (BRNNs) and empirically validate its efficacy, demonstrating that KDL excels in performance and interpretability by integrating label knowledge and learning logical rules. Particularly, BRNN with KDL outperforms the recent interpretable baselines with significant improvements of 3.57 %, 3.09 %, 2.66 %, 1.15 % and 1.19 % in terms of macro F₁ metric on GEVIA dataset. Furthermore, we conduct case studies to illustrate how KDL generates explanations consistent with the underlying decision-making logic of deep models through causal reasoning. Our assessment showcases that KDL offers explanations consistent with the decision logic of these models. We also uncover the concepts deep models rely on for decision-making and investigate how KDL-based BRNNs aid users by conducting a comparative experiment involving humans.