TigreGotico/EAT
收藏资源简介:
EAT(预期答案类型)数据集是一个高质量的问题分类数据集,基于TREC问题分类法,并增加了现代类别和严格的预期答案类型(EAT)验证。该数据集旨在训练和评估NLP模型,根据问题的预期答案的语义类别而非表面关键词进行分类。与原始TREC数据集不同,此版本包括:1. **BOOL:yesno**类别用于极性问题的分类;2. 每个样本的**已验证答案列**以确保标签准确性;3. **结构多样性**(直接、情态、命令式和上下文表达);4. **长度分层**(短、中、长和超长样本的平衡分布)。数据集通过大型语言模型(LLM)在人类数据工程师的严格监督下生成,并采用“生成-验证-清理”循环确保标签准确性。数据集以TSV格式提供,包含标签、问题、答案和语言代码等列。分类分类法包括主要标签(如ABBR、BOOL、DESC等)和层次化二级标签。生成过程中遵循六项基本原则,确保样本的多样性和质量。
The **EAT (Expected Answer Type)** dataset is designed to train and evaluate NLP models in the task of classifying questions not by their surface keywords, but by the semantic category of their expected answer. Unlike original TREC datasets, this version includes: 1. **BOOL:yesno** category for polar questions; 2. **Verified Answer Column** for every sample to ensure grounding and label accuracy; 3. **Structural Diversity** (Direct, Modal, Imperative, and Contextual phrasings); 4. **Length Stratification** (Balanced distribution of Short, Medium, Long, and Very Long samples). The dataset is synthetically generated using Large Language Models (LLM) under the strict supervision of a human data engineer, with an iterative "Generate-Validate-Clean" loop to ensure label accuracy. It is provided in TSV format with columns for label, question, answer, and language code. The classification taxonomy includes main labels (e.g., ABBR, BOOL, DESC) and hierarchical secondary labels. Generation follows six fundamental mandates to ensure diversity and quality.




