Introduction
Detecting harmful behavior on social media has evolved beyond simple keyword matching. Modern moderation requires distinguishing between nuanced, often overlapping behaviors such as sarcasm, teasing, and trolling. PromptAug is a novel data augmentation framework that leverages Large Language Models (LLMs) to generate high-quality training data for these fine-grained classifications. By combining a structured four-part prompting scheme with a rigorous three-step filtering process, PromptAug addresses the critical challenges of data scarcity and annotation complexity.
The Challenge of Fine-Grained Conflict
Building robust models for social media conflict detection is hindered by three primary factors: data scarcity due to API restrictions, annotation complexity that leads to human inconsistency, and LLM guardrails that often prevent the generation of the "offensive" text necessary for training detection systems.
The researchers focused on a specialized "Conflict Dataset" derived from a 16-month study of Facebook brand communities, featuring six distinct classes of behavior ranging from harmless banter to abusive threats.
The PromptAug Framework
PromptAug introduces a two-step methodology designed to generate diverse, label-consistent data while navigating LLM safety constraints.
1. Structured Prompting
Instead of generic requests, PromptAug uses a structured approach to guide the LLM:
- Instruction: Enforces a specific format (e.g., numbered lists) to prevent rambling.
- Context: Sets the scene (e.g., "directed at other social media users").
- Examples: Employs few-shot learning to tether the LLM to existing class boundaries.
- Definition: Provides explicit dictionary-style definitions for each behavior (e.g., defining Sarcasm as "humorous communication in a cynical tone").
2. Rigorous Filtering
To ensure data quality, every generated example must pass through three "Binary Prompt Assertions":
- Label Assertion: Confirms the example actually belongs to the intended class.
- Characteristic Assertion: Verifies the presence of specific behavioral traits defined in the prompt.
- Context Assertion: Ensures the output is a social media comment rather than the LLM providing advice or meta-commentary.
Experimental Results
The framework was evaluated using a BERT classifier on the Conflict dataset and the Crowdflower Emotion benchmark.
- Performance Gains: PromptAug achieved a statistically significant 2% increase in Accuracy and F1-score over models trained on original data alone, outperforming other augmentation methods like EDA, CBERT, and AugGPT.
- Low-Data Scenarios: The method proved most powerful in "cold start" situations. When only 20% of training data was available, PromptAug improved accuracy by 13% and F1-score by 16% over the baseline.
- Robustness: Improvements were consistent across various architectures, including CNNs and DistilBERT.
Discussion and Future Work
PromptAug demonstrates that structured definitions and context can effectively bypass LLM "refusal" behaviors while maintaining high label fidelity. However, the authors note that linguistic fluidity—where boundaries between classes like "Trolling" and "Teasing" are naturally blurry—remains a challenge. Future work will investigate whether larger models (like GPT-4) offer better results and how to ensure synthetic data does not reinforce social biases. Integrating human-in-the-loop reviews into the filtering stage could further refine the "Characteristic Assertion" check.
Summary updated