Relation extraction methods are currently dominated by deep neural models, which capture complex statistical patterns while being brittle and vulnerable to perturbations in data and distribution. Explainability techniques offer a means for understanding such vulnerabilities, and thus represent an opportunity to mitigate future errors; yet, existing methods are limited to describing what the model 'knows', while totally failing at explaining what the model does not know. This paper presents a new method for diagnosing model predictions and detecting potential inaccuracies. Our approach involves breaking down the problem into two components: (i) determining the necessary knowledge the model should possess for accurate prediction, through human annotations, and (ii) assessing the actual knowledge possessed by the model, using explainable AI methods (XAI). We apply our method to several relation extraction tasks and conduct an empirical study leveraging human specifications of what a model should know and does not know. Results show that human workers are capable of accurately specifying the model should-knows, despite variations in the specification, that the alignment between what a model really knows and what it should know is indeed indicative of model accuracy, and that the unknowns identified through our methods allow to foresee future errors that may or may not have been observed otherwise.

Noisy labels represent one of the key issues in supervised machine learning. Existing work for label noise reduction mainly takes a probabilistic approach that infers true labels from data distributions in low-level feature spaces. Such an approach is not only limited by its capability to learn high-quality data representations, but also by the low predictive power of data distributions in inferring true classes. To address those problems, we introduce Nessy, a neuro-symbolic system that integrates deep probabilistic modeling and symbolic knowledge for label noise reduction. Our deep probabilistic model infers the true classes of data instances with noisy labels by exploiting data distributions in an underlying latent feature representation space. For data instances where inference is not reliable enough, Nessy extracts symbolic rules and ranks them according to several utility metrics. Top-ranking rules are injected into the deep probabilistic model via expectation regularization, i.e., via a posterior regularization term constraining the class distribution in the objective function. In a real deployment over multiple relation extraction tasks, we demonstrate that Nessy is able to significantly improve the state of the art, by 7% accuracy and 10.7% AUC on average.

This paper presents Scalpel-CD, a first-of-its-kind system that leverages both human and machine intelligence to debug noisy labels from the training data of machine learning systems. Our system identifies potentially wrong labels using a deep probabilistic model, which is able to infer the latent class of a high-dimensional data instance by exploiting data distributions in the underlying latent feature space. To minimize crowd efforts, it employs a data sampler which selects data instances that would benefit the most from being inspected by the crowd. The manually verified labels are then propagated to similar data instances in the original training data by exploiting the underlying data structure, thus scaling out the contribution from the crowd. Scalpel-CD is designed with a set of algorithmic solutions to automatically search for the optimal configurations for different types of training data, in terms of the underlying data structure, noise ratio, and noise types (random vs. structural). In a real deployment on multiple machine learning tasks, we demonstrate that Scalpel-CD is able to improve label quality by 12.9% with only 2.8% instances inspected by the crowd.