Data Augmentation for Physical Commonsense Reasoning

Jiaqi Xu, Yukun Yang, Tianao Chen, Chuying Han

Keywords: NLP, Sense Reasoning

Abstract: Tiered Reasoning for Intuitive Physics (TRIP) is a new NLP dataset from the SLED Lab at University of Michigan. It poses a multi-layered physical commonsense reasoning task where AI systems must determine which of two stories (each describing a series of physical actions applied to household objects) is plausible. Further, they must justify their decision by identifying which sentences in the implausible story are conflicting with each other, and the specific physical state changes that cause the conflict. When we evaluate systems on this tiered prediction, baseline results are very low (only up to 10% of stories are classified correctly and verified with coherent supporting evidence). For this project, we are going to use three potential approaches to improve the baseline performance on TRIP.

Archival Option: Yes

Submission Number: 4