Complex-TR

Complex-TR数据集由阿里巴巴达摩院提出，专注于多答案和多跳跃时间推理，旨在提升大型语言模型在时间推理任务上的复杂性和鲁棒性。该数据集通过从Wikidata知识库中提取具有时间限定符的知识三元组构建，涵盖了从2010年到2020年的事实，特别关注时间-事件（L2）和事件-事件（L3）推理，这些类型要求将事件与时间轴对齐，比时间-时间（L1）推理更具挑战性。数据集的应用领域主要在于评估和提升大型语言模型在处理时间相关问题时的性能和准确性，特别是在处理多答案和多跳跃时间推理问题上的能力。The Complex-TR dataset was proposed by Alibaba DAMO Academy, focusing on multi-answer and multi-hop temporal reasoning, aiming to enhance the complexity and robustness of large language models (LLMs) in temporal reasoning tasks. This dataset is constructed by extracting knowledge triples with temporal qualifiers from the Wikidata knowledge base, covering factual knowledge from 2010 to 2020. It specifically focuses on temporal-event (L2) and event-event (L3) reasoning, which require aligning events with the timeline and are more challenging than time-time (L1) reasoning. The primary application scenarios of this dataset are to evaluate and improve the performance and accuracy of large language models when handling time-related problems, particularly their capabilities in addressing multi-answer and multi-hop temporal reasoning tasks.