Triple-Wise Perspective Knowledge Learning for Intelligent Machine Commonsense Reasoning

Enabling machines to possess commonsense reasoning within the intelligent Internet of Things (IoT) ecosystem plays a pivotal role in their capacity to attain autonomous decision-making and execution in complex tasks. The Winograd Schema Challenge (WSC) is a fundamental and challenging task in the field of commonsense reasoning, whose pairwise mutually exclusive properties and lack of contextual cues require machines with strong intelligence and skillful inference. Previous works that relied on pretrained models exhibited limited comprehension in commonsense because of overlooking the special property of the WSC task and solely relying on linguistic tendencies to learn superficial cues, failing to grasp the potential commonsense embedded within sentences. To address this issue, we propose a novel triple-wise perspective knowledge learning (TPKL) model for commonsense reasoning. Specifically, we introduce a new paradigm for addressing the WSC task by employing triplets instead of the conventional single or contrastive sentence inputs, which enables better compatibility with pairwise mutually exclusive features in WSC tasks. Additionally, we propose a triple-wise perspective that leverages anchor, positive, and negative sentences in a triplet construction to enable the model to comprehensively learn the pairwise mutually exclusive sentences, which can capture and utilize commonsense knowledge to distinguish between the various word senses under consideration. Extensive experiments conducted on three benchmark data sets demonstrate the superiority of our model over state-of-the-art baselines, improving PDP-60, WSC, and KnowRef benchmark with 3.3%, 4.4%, and 8.1%, respectively.