Finding Emotion in Image Descriptions: Crowdsourced Data

This dataset contains 660 images, each annotated with descriptions and mood labels. The images were originally created by users of the WordsEye text-to-scene system (https://www.wordseye.com/) and were downloaded from the WordsEye gallery. For each image, we used Amazon Mechanical Turk to obtain: (a) a literal description that could function as a caption for the image, (b) the most relevant mood for the picture (happiness, sadness, anger, surprise, fear, or disgust), (c) a short explanation of why that mood was selected. We published three AMT HITs for each picture, for a total of 1980 captions, mood labels, and explanations. This data was used for the machine learning experiments presented in: Morgan Ulinski, Victor Soto, and Julia Hirschberg. Finding Emotion in Image Descriptions. In Proceedings of the First International Workshop on Issues of Sentiment Discovery and Opinion Mining, WISDOM '12, pages 8:1-8:7. Please cite this paper if you use this data.

本数据集包含660张图像，每张均标注有描述文本与情绪标签。所有图像最初由WordsEye文本转场景系统（WordsEye text-to-scene system）的用户生成，并从WordsEye图库中下载。针对每张图像，我们通过亚马逊机械Turk（Amazon Mechanical Turk）平台采集三类标注信息：(a) 可作为该图像配图标题的直白描述文本；(b) 该图片最契合的情绪标签（包括快乐、悲伤、愤怒、惊讶、恐惧或厌恶）；(c) 选择对应情绪标签的简短解释说明。我们为每张图片发布3个亚马逊机械Turk人类智能任务（Human Intelligence Task，简称HIT），最终累计获得1980条标题、情绪标签及解释文本。该数据集被用于Morgan Ulinski、Victor Soto与Julia Hirschberg开展的机器学习实验研究，相关论文为《Finding Emotion in Image Descriptions》，收录于第一届情感发现与观点挖掘国际研讨会（International Workshop on Issues of Sentiment Discovery and Opinion Mining，WISDOM '12）论文集，页码为8:1-8:7。若使用本数据集，请引用该论文。