Explainable Methods for Information Extraction

Abstract

Explainability, the ability to understand and explain the behavior of a model, becomes especially important if human users are to trust machine learning systems. In the natural language processing domain, traditional methods, such as rule-based methods, have offered high degrees of interpretability; one could simply look at the rule to understand, to a reasonable degree, how (and why) a specific decision is being made. Yet, these methods are limited in their expressivity, leading to a shift towards the more capable deep-learning systems. In this dissertation, we revisit rule-based methods, but this time equipped with the recent advances in the deep learning field. At a high-level, we explore neuro-symbolic approaches that combine traditional rule-based systems, which are explainable but not expressive, with neural networks, which have high representation power but are opaque. To this end, we explore three directions to improving the performance of rule-based systems while maintaining their inherent advantages: explainability and pliability. First, we explore how a neural network itself could be used to generate the rules (generation), thus easing the amount of linguistic expertise a domain expert needs to use rule-based systems. Second, we explore how to select which rules should be used (filtering), by leveraging graph-based algorithms. Third, we explore how neural networks can augment the application of such rules over text (applying), allowing for degrees of matching rather than strict binary decisions.