Which side are you on? Insider-Outsider classification in conspiracy-theoretic social media

Pavan Holur, Tianyi Wang, Shadi Shahsavari, Timothy R. Tangherlini, Vwani P. RoychowdhuryACL 2022 · DOI

Introduction

Conspiracy theories on social media often revolve around a stark "us vs. them" worldview. Identifying the actors in these narratives—who is an ally (an Insider) and who is a threat (an Outsider)—is crucial for understanding the spread of misinformation. However, traditional sentiment analysis often fails to capture these dynamics because group identity is highly contextual and transitive.

The Challenge of Context-Sensitivity

Standard NLP tools struggle with the nuances of conspiracy rhetoric. For instance, a phrase like "the government is providing vaccines" might be seen as neutral or positive by a general-purpose model. Yet, in a vaccine-hesitant community, this frames "the government" as an Outsider.

Context-dependent labeling: The same entities can be classified as Insiders or Outsiders based on their narrative role.

As shown in the figure above, the identity of an agent is tied to the actions they perform (e.g., "kill" vs "save") and the entities they interact with. This context-sensitivity, combined with the fragmented and non-standard language of fringe platforms, necessitates a specialized approach that prioritizes narrative roles over simple keyword sentiment.

Methods: The NP2IO Architecture

To address these challenges, researchers developed NP2IO (Noun-Phrase-to-Insider-Outsider). This model moves beyond simple keyword matching to focus on the syntactic and contextual relationships within a sentence.

The NP2IO model architecture, utilizing spaCy for entity detection and a transformer backbone for token classification.

The pipeline begins by using spaCy to detect noun chunks, which represent the potential actors in the narrative. These chunks are then processed by a transformer-based model (such as BERT or RoBERTa) with a token-classification head. By training on the relationships between entities (e.g., "X causes Y"), the model learns to infer group membership even for entities it has never encountered before.

The CT5K Dataset

The study introduced CT5K, a dataset of 5,000 samples from platforms like Reddit, 4chan, and 8kun. Unlike standard sentiment datasets for product reviews (like laptop or restaurant reviews), CT5K captures the unique linguistic patterns of conspiracy theories. Experiments showed that models trained on traditional sentiment data perform poorly on CT5K, confirming that Insider-Outsider classification is a distinct task from standard sentiment analysis.

Key Results and Insights

NP2IO achieved a classification accuracy of 65%, significantly outperforming baseline models like Naive Bayes (~52%). Most notably, the model demonstrated strong zero-shot generalization, maintaining 69% accuracy on noun phrases not present in the training set. This suggests the model is truly learning the "language of threat" rather than just memorizing a list of known villains.

Furthermore, adversarial testing proved the model's robustness. When presented with "pro-microchip" sentences, the model was able to flip its classification of "microchips" from Outsider to Insider 80% of the time, demonstrating that it prioritizes immediate context over pre-existing biases in the training data.

Conclusion

NP2IO represents a significant step forward in automated narrative analysis. By focusing on roles and relationships rather than just emotion, it provides a more nuanced tool for researchers and moderators to map the shifting landscapes of online conspiracy theories. Future work will aim to expand this approach to other domains, such as political polarization and climate change denial, while improving the transparency of the model's causal reasoning.

Summary updated

BibTeX
@inproceedings{holur2022which,
  author = {Pavan Holur and Tianyi Wang and Shadi Shahsavari and Timothy R. Tangherlini and Vwani P. Roychowdhury},
  title = {{Which side are you on? Insider-Outsider classification in conspiracy-theoretic social media}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2022},
  doi = {10.18653/v1/2022.acl-long.341},
}
Figure