Dataset Construction for Fine-Grained Emotion Analysis in Catering Review Data

Emotion datasets serve as the foundation for training and evaluating emotion analysis models. By analyzing emotion datasets, we can gain users’ emotional tendencies and feedback in specific field. In this paper, we construct a large-scale fine-grained emotion classification dataset in catering field, named CateringEmo17, to enhance service quality and improve user experience. We employed an emotion annotation method that combines semi-automatic annotation and manual annotation to label 44,158 reviews of catering field with 17 different emotions.In the experiments we first trained a BERT-based model on the CateringEmo17 dataset to verify the accuracy of fine-grained emotion classification. Them we mapped fine-grained emotions to coarse-grained emotion to test the accuracy of coarse-grained emotion classification through a transformation experiment. Finally, we compared the model trained on a general emotion dataset, GoEmotion, with the same dataset for classification results. In the dataset anaysis, we validated the fine-grained emotion model on two restaurant dimensions including type, location. Experimental results indicate that our proposed fine-grained emotion classification model can provide potential information in the catering field, while traditional coarse-grained emotion classification models cannot capture this information.