IMPERFECTIVENLI
收藏资源简介:
IMPERFECTIVENLI是由慕尼黑大学和东京大学的研究团队创建的诊断性数据集,旨在探究大语言模型对未完成体悖论的理解能力。该数据集包含400个经过严格人工验证的例句,涵盖100个目标性动词和100个非目标性动词,通过模板化生成确保句法一致性。数据构建过程结合了Gemini辅助重写和人工标注,重点关注动词的体态分类及逻辑条件交叉分析。该数据集主要用于自然语言推理领域,揭示语言模型在事件语义理解中存在的目标性偏差问题,为改进模型对动作过程与结果状态的逻辑区分提供评估基础。
IMPERFECTIVENLI is a diagnostic dataset developed by research teams from Ludwig Maximilian University of Munich and the University of Tokyo, which aims to investigate the capacity of large language models (LLMs) to comprehend the imperfective paradox. This dataset comprises 400 rigorously human-validated example sentences, covering 100 telic verbs and 100 atelic verbs, and is generated via templatization to ensure syntactic consistency. The dataset construction process integrates Gemini-assisted rewriting and manual annotation, with a focus on verb aspect classification and cross-analysis of logical conditions. Primarily applied in the field of natural language inference (NLI), this dataset is designed to reveal the telic bias of language models in event semantic understanding, providing an evaluation foundation for enhancing the models' ability to logically differentiate between action processes and resultant states.
Imperfective Paradox 数据集概述
数据集简介
本仓库提供了一个用于评估大型语言模型在未完成体悖论上的简化框架,该评估基于自然语言推理任务。其旨在测试逻辑推理能力,特别聚焦于针对活动与成就两类动作的未完成体(判断一个动作是否已完成)。本仓库提供了一个诊断数据集 ImperfectiveNLI 以及用于进行实验的推理代码。
核心特性
- 多策略提示:无需更改代码即可在不同推理技术间轻松切换:
zero-shot:标准的零样本分类(真/假/未知)。dap:在系统提示中注入特定的语言学规则(活动动词与成就动词)。cot:关注动作时间端点的思维链方法。counterfactual:反事实方法,要求模型首先列出潜在的现实世界干扰因素,迫使模型思考干扰。
- 模型无关性:兼容任何 Hugging Face Transformer 模型(如 Llama 3、Mistral、Qwen、DeepSeek 等)。
- 鲁棒的格式化:自动为经过指令调优的模型应用正确的
chat_template。
数据格式
数据集以 JSON 格式组织,示例如下: json [ { "id": "A_001", "group": "A_Interrupted_Accomplishment", "verb_class": "Creation", "verb": "build", "premise": "The carpenter was building a gazebo, but a storm destroyed the frame before the roof was on.", "hypothesis": "The carpenter built a gazebo.", "label": "False", }, ... ]
每个数据条目包含唯一标识符 id、所属组别 group、动词类别 verb_class、具体动词 verb、前提句 premise、假设句 hypothesis 以及真实标签 label。
相关资源
- 论文链接:https://arxiv.org/abs/2601.09373

- 1The Imperfective Paradox in Large Language Models慕尼黑大学; 东京大学 · 2026年



