Semantic extensibility captures the semantic side of productivity. It is the likelihood that a given sense of a linguistic expression will support extension to new senses. Even though linguistic expressions are naturally polysemous, semantic extensibility is constrained. In previous literature, it has been argued that semantic extensions are motivated by mostly one-directional conceptual operations such as metaphor and metonymy, and that in any polysemous expression only one or a few so-called ‘sanctioning’ senses have privileged status in supporting new extensions. One factor believed to determine sanctioning status is high frequency. Drawing on three case studies from the history of English, involving change in the adjective awful, the preposition and adverb about and the multifunctional item so, this article provides diachronic evidence from semantic loss to support this view. On the one hand, it is shown that when old sanctioning senses go into decline, this also impacts the senses derived from them, underscoring the motivational relations that tie extended senses to sanctioning senses. On the other hand, what typically initiates a decline in a sanctioning sense is a frequency increase elsewhere in the polysemy network coincident with the emergence of a new sanctioning sense, underscoring the role of frequency in determining sanctioning status and the directionality of sanctioning relations.

Recent efforts in interpreting convolutional neural networks (CNNs) focus on translating the activation of CNN filters into a stratified Answer Set Program (ASP) rule-sets. The CNN filters are known to capture high-level image concepts, thus the predicates in the rule-set are mapped to the concept that their corresponding filter represents. Hence, the rule-set exemplifies the decision-making process of the CNN w.r.t the concepts that it learns for any image classification task. These rule-sets help understand the biases in CNNs, although correcting the biases remains a challenge. We introduce a neurosymbolic framework called NeSyBiCor for bias correction in a trained CNN. Given symbolic concepts, as ASP constraints, that the CNN is biased toward, we convert the concepts to their corresponding vector representations. Then, the CNN is retrained using our novel semantic similarity loss that pushes the filters away from (or toward) learning the desired/undesired concepts. The final ASP rule-set obtained after retraining, satisfies the constraints to a high degree, thus showing the revision in the knowledge of the CNN. We demonstrate that our NeSyBiCor framework successfully corrects the biases of CNNs trained with subsets of classes from the Places dataset while sacrificing minimal accuracy and improving interpretability.