Fine-tuning for multi-domain and multi-label uncivil language detection

Kadir Bulut Özler, Kate Kenski, Stephen A. Rains, Yotam Shmargad, Kevin Coe, Steven BethardWOAH 2020 · DOI

Introduction

Online incivility is a multifaceted problem. Detecting it effectively requires models that can navigate diverse domains—from local news comments to political tweets—and recognize overlapping categories of abuse, such as simultaneous vulgarity and threats. Researchers from the University of Arizona and the University of Utah address these challenges by evaluating BERT-based fine-tuning strategies for multi-domain and multi-label uncivil language detection.

Methodology

The researchers compared three fine-tuning strategies using BERT:

  • Single: Training a separate model for each specific domain or label.
  • Joint: Training one model on all domains or labels simultaneously. For multi-label tasks, this involves using independent sigmoid activations for each category.
  • Joint → Single: A staged approach where a model is first trained jointly and then further fine-tuned on specific domains or labels to specialize.

Key Insights

Complexity vs. Simplicity in Multi-Domain Learning

A major finding is that "less is more" when dealing with multiple domains. In multi-domain experiments, the Joint model—simply pooling all training data—often performed as well as or better than the more complex staged fine-tuning (Joint → Single). As shown in Figure 1, the Joint model achieved superior or competitive $F_1$ scores across local news and tweet datasets, suggesting that BERT is robust enough to learn core linguistic features of incivility that transcend specific platforms.

Performance metrics (Precision, Recall, F1) for multi-domain training strategies across three datasets.

The Power of Independence in Multi-Label Tasks

Counter-intuitively, for multi-label detection, independent binary classifiers (the "Single" approach) frequently outperformed joint multi-task models. While multi-task learning is often expected to benefit from shared hidden layers, these results suggest that for incivility, treating labels independently leads to better performance, particularly for rare or specific categories like threats or identity hate.

Results and Benchmarks

The researchers established new state-of-the-art (SOTA) performance across several benchmarks. For local news comments, they achieved an $F_1$ of 0.56 for name-calling, a significant improvement over the previous SOTA of 0.48. On the Wikipedia dataset, they reached an AUC of 0.990, matching or exceeding previous benchmarks while providing a more granular analysis of specific labels. Figure 2 details the performance across these specific categories, highlighting the effectiveness of the "Single" model approach for granular detection.

Multi-label classification performance (F1 and AUC) across different training strategies for Local News and Wikipedia comments.

Conclusion

This research demonstrates that simple, robust fine-tuning strategies can match or exceed the performance of more complex architectures for uncivil language detection. By providing the first benchmarks for several datasets and achieving new SOTA results on others, this work offers a clear path forward for building more effective moderation tools across the diverse landscape of the internet.

Summary updated

BibTeX
@inproceedings{ozler2020fine,
  author = {Kadir Bulut Özler and Kate Kenski and Stephen A. Rains and Yotam Shmargad and Kevin Coe and Steven Bethard},
  title = {{Fine-tuning for multi-domain and multi-label uncivil language detection}},
  booktitle = {WOAH},
  year = {2020},
  doi = {10.18653/v1/2020.alw-1.4},
}
Figure