Should We Afford Affordances? Injecting ConceptNet Knowledge into BERT-Based Models to Improve Commonsense Reasoning Ability

Andrzej Gretkowski, Dawid Wisniewski, Agnieszka Lawrynowicz

Abstract: Recent years have shown that deep learning models pre-trained on large text corpora using the language model objective can help solve various tasks requiring natural language understanding. However, many commonsense concepts are underrepresented in online resources because they are too obvious for most humans. To solve this problem, we propose the use of affordances – common-sense knowledge that can be injected into models to increase their ability to understand our world. We show that injecting ConceptNet knowledge into BERT-based models leads to an increase in evaluation scores measured on the PIQA dataset.

External IDs:dblp:conf/ekaw/GretkowskiWL22