TRIP (Tiered Reasoning for Intuitive Physics)

直觉物理学的分层推理 (TRIP) 是一种新颖的常识推理数据集，具有密集的注释，可以对机器的推理过程进行多层评估。 TRIP 作为物理常识推理的基准，为合理性预测的最终任务提供推理痕迹。该数据集由描述具体身体动作序列的人类创作故事组成。给定两个由单独合理的句子组成的故事，并且仅相差一个句子（即句子 5），建议的任务是确定哪个故事更合理。要理解这样的故事并做出这样的预测，就必须了解动词的因果关系和前提，以及直觉物理学的规则。
描述来自：直觉物理学的分层推理：迈向可验证的常识语言理解

Hierarchical Reasoning for Intuitive Physics (TRIP) is a novel commonsense reasoning dataset with dense annotations that enables multi-layer evaluation of machine reasoning processes. As a benchmark for physical commonsense reasoning, TRIP provides reasoning traces for the final task of plausibility prediction. This dataset consists of human-written stories describing sequences of concrete physical actions. Given two stories each composed of individually plausible sentences that differ in exactly one sentence (i.e., Sentence 5), the proposed task is to determine which story is more plausible. To understand such stories and make such predictions, one must grasp the causal relations and preconditions of verbs, as well as the rules of intuitive physics. Description sourced from: Hierarchical Reasoning for Intuitive Physics: Towards Verifiable Commonsense Language Understanding

搜集汇总

背景与挑战

背景概述

TRIP是一个用于直觉物理学分层推理的常识性数据集，通过人类创作的故事来评估机器的物理常识推理能力，要求比较两个故事的合理性。该数据集由密歇根大学于2021年发布，旨在支持可验证的常识语言理解研究。