SMARTER: A Data-efficient Framework to Improve Toxicity Detection with Explanation via Self-augmenting Large Language Models

TL;DR

SMARTER tackles toxicity detection with explanations in a data-efficient framework. It combines Stage 1 self-augmentation and preference-alignment (DPO/KTO) on few-shot data with Stage 2 cross-model refinement to align explanations across LLMs. Evaluations on HateXplain, Latent Hate, and Implicit Hate using Llama-3.1-8B-Instruct and COT-T5-XL show up to 13.5% macro-F1 gains with far less training data than full-data baselines, while maintaining explainability. The approach demonstrates scalable, explainable content moderation in low-resource settings, though benefits depend on model architecture and require human oversight to manage biases and quality of explanations.

Abstract

WARNING: This paper contains examples of offensive materials. To address the proliferation of toxic content on social media, we introduce SMARTER, we introduce SMARTER, a data-efficient two-stage framework for explainable content moderation using Large Language Models (LLMs). In Stage 1, we leverage LLMs' own outputs to generate synthetic explanations for both correct and incorrect labels, enabling alignment via preference optimization with minimal human supervision. In Stage 2, we refine explanation quality through cross-model training, allowing weaker models to align stylistically and semantically with stronger ones. Experiments on three benchmark tasks -- HateXplain, Latent Hate, and Implicit Hate -- demonstrate that SMARTER enables LLMs to achieve up to a 13.5% macro-F1 improvement over standard few-shot baselines while using only a fraction of the full training data. Our framework offers a scalable strategy for low-resource settings by harnessing LLMs' self-improving capabilities for both classification and explanation.

Figures (15)

• Figure 1: Bar plots for $K$-shot classification experiments on 3 datasets using Llama and T5 models. Macro F1 scores and percentage change over Baseline are displayed on top. Results for Baselines and DPO-augmented variants for $K \in \{16, 32, 64, 128\}$ are displayed on the left subfigures. Results for $K=256$ of Baseline, KTO, DPO-augmented and its other variants on various sub-sampling strategies (section \ref{['cls_pipeline']}) are shown on the right. Horizontal lines show the F1 scores for Full models that use all training data.
• Figure 2: Macro-F1 scores on test portion of the 3 test sets for T5 and Llama cross-model refinement regimen. In each figure: the first and last bars are scores reported with only T5 model at $K=128$ and $K=256$ (directly from \ref{['fig:cls_result']}); the middle bars are results after further finetuned using data from the complementary $K=128$ shots of counterpart model. Split-color bars: bottom color indicates the original model; top color indicates the counterpart model for additional cross-training.
• Figure 3: Template to collect annotation for preference on explanations for Amazon Mechanical Turks crowdworkers.
• Figure 4: Self-augmenting pipeline: for each post, explanations are conditionally generated for the gold label and all incorrect labels using prompt template in \ref{['fig:cls_prompt']}. For DPO, data consists of matching the explanation of the correct label with another incorrect label's.
• Figure 5: Prompt template for classification tasks without explanation generation. The model is instructed to directly output a label based on the provided definitions.