SarcNet: A Multilingual Multimodal Sarcasm Detection Dataset

SarcNet is a novel benchmark for multilingual and multimodal sarcasm detection, introduced at LREC-COLING 2024. It addresses the limitations of single-language datasets by providing 3,335 image-text pair samples in both English and Chinese.

Dataset Summary

In contrast to traditional datasets that use a single unified label, SarcNet employs a separated annotation schema. Each sample is distinctly labeled across three modalities, providing over 10,000 labels in total:

• Text Label: Sarcasm detected in the text alone.
• Image Label: Sarcasm detected in the image alone.
• Multimodal Label: Sarcasm detected through the combination of both modalities.

Data Splits

The dataset follows an official split of 60% Train, 20% Val, and 20% Test.

Language	Train	Validation	Test	Total
Chinese	1,242	424	399	2,065
English	756	247	267	1,270
Total	1,998	671	666	3,335

Dataset Features

The columns in this reformatted version are:

• id: The unique identifier (original image path).
• image: The decoded image file.
• text: The associated textual statement.
• text_label: ClassLabel (0: non-sarcastic, 1: sarcastic).
• image_label: ClassLabel (0: non-sarcastic, 1: sarcastic).
• multi_label: ClassLabel (0: non-sarcastic, 1: sarcastic).
• language: The detected language code (en or zh).

Usage

Because this dataset is configured with subsets, you can load your desired language directly:

from datasets import load_dataset

# Load English only
dataset_en = load_dataset("your-username/sarcnet-multimodal", "en")

# Load Chinese only
dataset_zh = load_dataset("your-username/sarcnet-multimodal", "zh")

Citation

If you use this dataset in your research, please cite the original work:

@inproceedings{yue-etal-2024-sarcnet, title = "{S}arc{N}et: A Multilingual Multimodal Sarcasm Detection Dataset", author = "Yue, Tan and Shi, Xuzhao and Mao, Rui and Hu, Zonghai and Cambria, Erik", booktitle = "Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024)", year = "2024", address = "Torino, Italia", pages = "14325--14335", publisher = "ELRA Language Resource Association", url = "https://aclanthology.org/2024.lrec-main.1248" }