Dataset for SemEval-2024 Task 3

The dataset ECF 2.0 for SemEval-2024 Task 3: Multimodal Emotion Cause Analysis in Conversations is released here.

In our preliminary work MECPE, we constructed the ECF (1.0) dataset. For this SemEval competition, the entire ECF (1.0) dataset serves as the training data, and we have additionally annotated a new test set as the evaluation data, together constituting the ECF 2.0 dataset.

File Description

The files Subtask_1_train.json and Subtask_2_train.json contain the training data for two subtasks in SemEval 2024, where all instances are stacked into a list and each of them is stored in a dictionary. The files Subtask_1_test.json and Subtask_2_test.json contain the evaluation data, using the same format.
For Subtask 2, you can refer to the JSON files to obtain timestamps and download the raw videos from the website to process the multimodal data. ⚠️ Due to potential copyright issues, we do not provide pre-segmented video clips.

📢 Note: To ensure the fairness of our competition, the evaluation data released during the evaluation phase of SemEval 2024 included some noise data that was not intended for evaluation. In the newly released version, the noise data has been removed, and the labels have been made publicly available.

Dataset Statistics

	Training Data	Evaluation data
# of conversations	1,374	341
# of utterances	13,619	3,101