From Symbolic to Neurosymbolic Information Extraction

Neural methods dominate the landscape of natural language processing today. However, they are not without limitations: they exhibit low explainability, are fragile, and, in the case of large language models, face reliability issues, often manifesting as “hallucinations.” These shortcomings constrain their applicability in critical domains such as healthcare, law, and intelligence. In contrast, symbolic approaches, such as rule-based systems, are explainable and pliable but lack the generalization capabilities of modern deep learning methods and are often labor-intensive to develop. In this chapter, we advocate for hybrid approaches that combine symbolic and neural methods to leverage the strengths of both. Specifically, to address the labor-intensive nature of rule creation, we introduce a novel strategy for rule synthesis that requires minimal human supervision. Furthermore, we propose a method to enhance the generalization capabilities of rules by semantically matching them to text, rather than relying on the encoded lexical-syntactic patterns. We evaluate these proposed approaches in the context of information extraction in biomedical and open domains and demonstrate that they outperform purely neural methods while maintaining the explainability and adaptability of symbolic systems.