| license: mit | |
| task_categories: | |
| - text-classification | |
| language: | |
| - en | |
| size_categories: | |
| - 100K<n<1M | |
| tags: | |
| - deception-detection | |
| - fake-news | |
| This repository contains the **SEPSIS** dataset, introduced in the paper [SEPSIS: I Can Catch Your Lies -- A New Paradigm for Deception Detection](https://huggingface.co/papers/2312.00292). | |
| The SEPSIS dataset is a novel, large-scale annotated dataset designed for deception detection, specifically focusing on "lies of omission" using Natural Language Processing (NLP) techniques. It comprises **876,784 samples**, curated by amalgamating a popular large-scale fake news dataset and scraped news headlines from the Twitter handle of the Times of India. | |
| Each sample in the dataset is labeled with four distinct layers of annotation, providing a comprehensive understanding of deceptive content: | |
| * **Type of omission**: Categorizes the type of omission into speculation, bias, distortion, sounds factual, and opinion. | |
| * **Colors of lies**: Identifies the moral or ethical implications of the lie (e.g., black, white). | |
| * **Intention of such lies**: Indicates the underlying purpose of the lie (e.g., to influence). | |
| * **Topic of lies**: Specifies the subject matter of the deceptive content (e.g., political, educational, religious). | |
| This dataset aims to encourage further research in the field of deception detection and explores the intricate relationship between lies of omission and propaganda techniques. |