Large-scale, pre-trained language models (LMs) have achieved human-level performance on a breadth of language understanding tasks. However, evaluations only based on end task performance shed little light on machines’ true ability in language understanding and reasoning. In this paper, we highlight the importance of evaluating the underlying reasoning process in addition to end performance. Toward this goal, we introduce Tiered Reasoning for Intuitive Physics (TRIP), a novel commonsense reasoning dataset with dense annotations that enable multi-tiered evaluation of machines’ reasoning process. Our empirical results show that while large LMs can achieve high end performance, they struggle to support their predictions with valid supporting evidence. The TRIP dataset and our baseline results will motivate verifiable evaluation of commonsense reasoning and facilitate future research toward developing better language understanding and reasoning models.
To enable collaboration and communication between humans and agents, this paper investigates learning to acquire commonsense evidence for action justification. In particular, we have developed an approach based on the generative Conditional Variational Autoencoder(CVAE) that models object relations/attributes of the world as latent variables and jointly learns a performer that predicts actions and an explainer that gathers commonsense evidence to justify the action. Our empirical results have shown that, compared to a typical attention-based model, CVAE achieves significantly higher performance in both action prediction and justification. A human subject study further shows that the commonsense evidence gathered by CVAE can be communicated to humans to achieve a significantly higher common ground between humans and agents.
Despite recent advances in knowledge representation, automated reasoning, and machine learning, artificial agents still lack the ability to understand basic action-effect relations regarding the physical world, for example, the action of cutting a cucumber most likely leads to the state where the cucumber is broken apart into smaller pieces. If artificial agents (e.g., robots) ever become our partners in joint tasks, it is critical to empower them with such action-effect understanding so that they can reason about the state of the world and plan for actions. Towards this goal, this paper introduces a new task on naive physical action-effect prediction, which addresses the relations between concrete actions (expressed in the form of verb-noun pairs) and their effects on the state of the physical world as depicted by images. We collected a dataset for this task and developed an approach that harnesses web image data through distant supervision to facilitate learning for action-effect prediction. Our empirical results have shown that web data can be used to complement a small number of seed examples (e.g., three examples for each action) for model learning. This opens up possibilities for agents to learn physical action-effect relations for tasks at hand through communication with humans with a few examples.