(Comet-)Atomic2020subscriptsuperscriptabsent2020{{}^{20}_{20}}: On Symbolic and Neural Commonsense Knowledge Graphs
Jena D. Hwang1, Chandra Bhagavatula, Ronan Le Bras1, Jeff Da1, Keisuke Sakaguchi1, Antoine Bosselut13 and Yejin Choi12 The authors contributed equally to this work.
Abstract
Recent years have brought about a renewed interest in commonsense representation and reasoning in the field of natural language understanding. The development of new commonsense knowledge graphs (CSKG) has been central to these advances as their diverse facts can be used and referenced by machine learning models for tackling new and challenging tasks. At the same time, there remain questions about the quality and coverage of these resources due to the massive scale required to comprehensively encompass general commonsense knowledge.
中文速览
手工构建的常识知识图谱（Commonsense Knowledge Graph, CSKG）永远无法做到面面俱到，而大型语言模型虽能记住大量事实，却在隐性常识推理上仍有明显短板——这正是本文要解决的核心矛盾。为此，作者构建了 Atomic²⁰₂₀，一个涵盖 23 种常识关系、133 万条日常推理知识的高质量知识图谱，专门收录了语言模型难以自行习得的社会、物理和事件类常识，并通过大规模人工评估，系统比较了它与 ConceptNet、原版 Atomic、TransOMCS 等主流资源在覆盖率和准确率上的差异。在此基础上，作者沿用 COMET 框架，将 Atomic²⁰₂₀ 用于微调预训练语言模型，训练出可按需生成新实体和新事件常识的神经知识模型（COMET-Atomic²⁰₂₀）。人工评估结果显示，这个基于 BART 的知识模型尽管参数量仅为 GPT-3 的约 1/430，却在常识生成任务上比 GPT-3 的少样本表现高出约 12 个百分点，有力证明了高质量符号知识对于弥补语言模型常识盲区的不可替代价值。
原文 arXiv:2010.05953;中英对照 + 大白话阅读 https://aha.fim.ai/paper/2010.05953v2