SCCD: A Session-based Dataset for Chinese Cyberbullying Detection

TL;DR

The paper addresses the lack of Chinese session-based cyberbullying data by introducing SCCD, a large, annotated dataset sourced from Weibo that provides fine-grained comment labels and session-level severity. It presents a two-stage annotation pipeline combining LLM-assisted labeling with human verification to produce reliable comment annotations and session classifications. Systematic experiments show that while transformer models and GPT-4 outperform baseline methods, detecting cyberbullying in Chinese remains challenging due to subtlety and contextual dependence, and comments within sessions substantially enhance performance. The dataset and analysis offer a valuable benchmark for advancing Chinese cyberbullying detection and enable more explainable, context-aware methods with potential benefits for online safety and moderation.

Abstract

The rampant spread of cyberbullying content poses a growing threat to societal well-being. However, research on cyberbullying detection in Chinese remains underdeveloped, primarily due to the lack of comprehensive and reliable datasets. Notably, no existing Chinese dataset is specifically tailored for cyberbullying detection. Moreover, while comments play a crucial role within sessions, current session-based datasets often lack detailed, fine-grained annotations at the comment level. To address these limitations, we present a novel Chinese cyber-bullying dataset, termed SCCD, which consists of 677 session-level samples sourced from a major social media platform Weibo. Moreover, each comment within the sessions is annotated with fine-grained labels rather than conventional binary class labels. Empirically, we evaluate the performance of various baseline methods on SCCD, highlighting the challenges for effective Chinese cyberbullying detection.

Figures (9)

• Figure 1: An illustration of a cyberbullying session from Weibo with repetitive offensive behaviors.
• Figure 2: Overview of data construction. The data annotation process involves two steps: first, using a large language model (LLM) to annotate the comments, and then manually labeling the conversations.
• Figure 3: The annotation guideline of comments. It outlines four key questions that determine the five labels.
• Figure 4: A sample of post information in the dataset.
• Figure 5: An example of comment information in the dataset.