Modeling Framing in Immigration Discourse on Social Media

Julia Mendelsohn, Ceren Budak, David Jurgens

Abstract

Julia Mendelsohn, Ceren Budak, David Jurgens. Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies.

2021.

Introduction

Framing is a strategic communication tool used to highlight specific aspects of an issue to influence public perception. While the framing strategies of political elites and news media have been extensively studied, the framing of complex issues by the general public on social media remains less understood. Traditional Natural Language Processing (NLP) approaches often rely on "issue-generic" typologies—broad categories like Economics or Legality. However, these categories can be too coarse to capture the nuances of polarized debates. For instance, a tweet discussing immigrants as economic contributors and one discussing them as job threats both fall under an "Economic" frame, despite representing opposing viewpoints. This research bridges political communication theory and NLP to provide a multi-dimensional analysis of immigration discourse on Twitter.

Methodology

The researchers developed a supervised learning framework to detect three distinct framing typologies:

  1. Issue-Generic Policy Frames: Broad categories such as Morality, Security, and Quality of Life.
  2. Immigration-Specific Frames: Roles assigned to immigrants, categorized as Victim, Hero, or Threat.
  3. Narrative Frames: Classification of tweets as Episodic (focused on specific events/individuals) or Thematic (focused on broad social trends).

The study utilized a dataset of 2.66 million English-language tweets from 2018-2019 across the US, UK, and EU. A novel gold-standard dataset of 4,500 manually labeled tweets was used to fine-tune a RoBERTa model for multi-label classification. Political ideology for US users was estimated using a Bayesian spatial following model.

Results

Model Performance and Hegemonic Framing

The fine-tuned RoBERTa model achieved an F1 score of 0.657, significantly outperforming baseline models. Notably, the researchers observed that the model performed consistently better at identifying frames in conservative tweets than in liberal ones across all typologies. This suggests that conservative discourse on immigration is more linguistically consistent and "hegemonic" compared to the more varied liberal discourse.

Predictive performance (F1 score) for Liberal vs. Conservative tweets across different frame types.

Ideological Stratification and the Masking Effect

The analysis revealed a clear ideological map of framing. While generic frames like "Economic" appeared only slightly conservative, the immigration-specific frames revealed a deep divide: liberals frequently used the Hero: Worker frame, while conservatives favored the Threat: Jobs frame. This confirms that generic frames can obscure the most contentious elements of public debate.

Association between specific frames and political ideology, showing how immigration-specific roles reveal deeper divides than generic policy frames.

Regional and Narrative Differences

Regional analysis showed that US-based discourse is more likely to utilize "extreme" frames like Crime & Punishment or Security & Defense compared to the UK or EU. Furthermore, liberals showed a preference for Episodic narrative frames (sharing individual stories), while conservatives leaned toward Thematic frames (discussing broad societal impacts).

Discussion

The study demonstrates that a multi-dimensional approach to framing is essential for capturing the complexity of social media discourse. The findings suggest that audience engagement is driven by different framing strategies: human-interest frames drive "Likes" (alignment), while political strategy frames drive "Retweets" (amplification). While these insights offer a powerful tool for understanding public opinion, the authors caution that such findings could be weaponized to spread harmful content more effectively. Future research will explore applying this multi-typology framework to other polarized issues such as climate change.

Summary updated

BibTeX
@inproceedings{mendelsohn2021modeling,
  author = {Julia Mendelsohn and Ceren Budak and David Jurgens},
  title = {{Modeling Framing in Immigration Discourse on Social Media}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2021},
  url = {https://openreview.net/forum?id=4xgTCGQLQbS},
}
Figure