Cross-community Adapter Learning (CAL) to Understand the Evolving Meanings of Norm Violation

Thiago Freitas dos Santos, Stephen Cranefield, Bastin Tony Roy Savarimuthu, Nardine Osman, Marco SchorlemmerIJCAI 2023 · DOI

TL;DR

This paper introduces Cross-community Adapter Learning (CAL), a novel framework that uses adapters within transformer models to learn and understand how the meaning of norm violations evolves across different online communities. CAL addresses the challenge of adapting to changing community norms in low-resource settings by enabling continuous learning and providing interpretable explanations for detected violations.

What Problem Does This Paper Address?

Online communities rely on norms to maintain order and foster inclusivity. However, the interpretation and enforcement of these norms, especially regarding violations like hate speech, are dynamic and can vary significantly between communities. This poses a challenge for automated systems designed to detect such violations, particularly in new or low-resource communities where labeled data is scarce. Existing approaches often struggle to adapt to evolving community views or provide clear explanations for their decisions. This paper tackles the problem of building adaptable and interpretable systems that can learn the nuanced meanings of norm violations across diverse online communities.

Methods

The core of the proposed solution is the Cross-community Adapter Learning (CAL) framework. CAL leverages pre-trained language models (PLMs) like DistilBERT, which are powerful for natural language understanding. The novelty lies in incorporating adapters – small, efficient neural network modules – between the layers of the PLM. These adapters are fine-tuned on community-specific data, allowing the model to adapt to new norms without retraining the entire PLM, thus addressing catastrophic forgetting and improving efficiency.

To understand why the model makes certain classifications, CAL employs the Integrated Gradients (IG) algorithm. IG provides a score indicating the relevance of each word in a sentence to the model's prediction. By analyzing these scores, CAL can explain how different communities interpret norm violations and how these interpretations evolve over time. The framework also incorporates incremental learning to process incoming data sequentially, allowing for continuous adaptation to emerging violation classes and shifting community views.

Datasets

The study uses data from Wikipedia for its target task, specifically focusing on the norm prohibiting hate speech. For source communities, the authors utilize three publicly available datasets:

  • Slack community: Provides "Swear" instances.
  • ConvAbuse dataset: Includes "Insult and Ableism" and "Sexual Harassment" instances.
  • DynamicGenerated dataset: Contains "Racism," "LGBTQIA+ Attack," and "Misogyny" instances.

The authors note that they define violation classes based on community member feedback, rather than relying on pre-established categories.

Key Insights and Ideas

  • Adapters are key for efficient adaptation: Fine-tuning only small adapter modules is significantly more efficient than fine-tuning entire PLMs, especially in low-resource scenarios.
  • Cross-community learning boosts performance: Training on data from multiple source communities before fine-tuning on the target community leads to better performance in detecting norm violations.
  • Interpretability reveals evolving norms: Integrated Gradients effectively highlight how the relevance of specific words for norm violation detection changes as a community's views evolve. This allows for a deeper understanding of the "why" behind violations.
  • Norms are fluid and context-dependent: The study demonstrates that what constitutes a norm violation is not static but is shaped by the specific community and its evolving discourse.

Key Results

The experiments show that CAL, when fine-tuned with cross-community data, significantly outperforms models trained solely on the target community data for most hate speech classes. For instance, the "Source - Target" setting (training on source, testing on target without fine-tuning) yielded the lowest performance, while "Target - Target" (fine-tuning a cross-community model on the target) showed the best results. The interpretability analysis revealed specific word shifts in relevance scores, illustrating how the model adapts its understanding of violations (e.g., "gay" becoming more relevant for LGBTQIA+ Attack detection in the target community, while words like "dick" and "fuck" lose relevance).

Limitations and Future Work

The authors acknowledge that for some classes, like "Sexual Harassment," the "Only Target" approach performed very well, suggesting that adding data might not always be beneficial if the target community's data is already highly representative. The precision for "Racism" was affected by the small number of instances and differences in the groups targeted by racism in source versus target communities.

Future work includes:

  • Conducting user studies to validate the usefulness of the model's interpretations for understanding community norms.
  • Exploring alternative adapter architectures to further improve training efficiency in low-resource settings.
  • Integrating CAL into non-ML solutions for monitoring norm violations in normative multi-agent systems.

Summary updated

BibTeX
@inproceedings{santos2023cross,
  author = {Thiago Freitas dos Santos and Stephen Cranefield and Bastin Tony Roy Savarimuthu and Nardine Osman and Marco Schorlemmer},
  title = {{Cross-community Adapter Learning (CAL) to Understand the Evolving Meanings of Norm Violation}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2023},
  doi = {10.24963/ijcai.2023/13},
}
Figure