L3Cube-MahaEmotions
收藏资源简介:
L3Cube-MahaEmotions 是一个高质量的马拉地语情感识别数据集,包含 11 种细粒度的情感标签。该数据集的训练数据使用大型语言模型 (LLMs) 进行合成标注,而验证集和测试集则由人工标注,以确保可靠的金标准基准。基于 MahaSent 数据集,我们应用了链式翻译 (CoTR) 提示技术,将马拉地语文本翻译成英语,并通过单个提示进行情感标注。GPT-4 和 Llama3-405B 被评估,由于标签质量优越,GPT-4 被选中用于训练数据标注。我们使用标准指标评估模型性能,并探索标签聚合策略(例如,并集、交集)。
L3Cube-MahaEmotions is a high-quality Marathi emotion recognition dataset containing 11 fine-grained emotion labels. Its training data was synthetically annotated using large language models (LLMs), while its validation and test sets were manually annotated to establish a reliable gold-standard benchmark. Built on the MahaSent dataset, we utilized the Chain-of-Translation (CoTR) prompting technique to translate Marathi text into English and conduct emotion annotation with a single prompt. Both GPT-4 and Llama3-405B were evaluated, and GPT-4 was ultimately selected for training data annotation due to its superior label quality. We adopted standard metrics to assess model performance and explored label aggregation strategies such as union and intersection.




