GRAB
收藏资源简介:
GRAB是一个用于评估金融风险分类的无监督主题模型的公共基准。它包含来自8247份文件的161.38万个句子,并且没有手动注释,而是通过结合FinBERT标记注意力、YAKE关键词信号和分类意识搭配匹配产生的跨度接地句子标签。标签锚定在风险分类法中,将193个术语映射到五个宏观类别下的21个细粒度类型;21个类型指导弱监督,而评估则是在宏观层面报告的。GRAB通过固定的数据集分割和鲁棒的指标(准确性、宏观F1、主题BERT分数和基于熵的有效主题数)统一了评估。该数据集、标签和代码使经典、基于嵌入、神经和混合主题模型在金融风险披露方面具有可重复性和标准化的比较。
GRAB is a public benchmark for evaluating unsupervised topic models for financial risk classification. It contains 1.6138 million sentences from 8,247 documents. Without manual annotations, its span-grounded sentence labels are generated by combining FinBERT token attention, YAKE keyword signals, and classification-aware collocation matching. The labels are anchored in a risk taxonomy, mapping 193 terms to 21 fine-grained types under five macro-categories; the 21 types are used to guide weak supervision, while evaluations are reported at the macro level. GRAB unifies model evaluation via fixed dataset splits and robust metrics including accuracy, macro-F1, TopicBERT score, and entropy-based effective topic number. This dataset, its accompanying labels and code, enable reproducible and standardized comparison of classical, embedding-based, neural, and hybrid topic models for financial risk disclosure.




