Lightnews — Scholar-powered news

Patrick Gerard @patrikgerard.bsky.social · Jun 22

⬆️ This work builds on our recent paper discussed in this thread, in which we build out cross-platform user networks by connecting users via shared ideas!

bsky.app/profile/patr...

Patrick Gerard @patrikgerard.bsky.social · Jun 22

While I'm here, I might as well post about my newest work with @luceriluc.bsky.social and @emilioferrara.bsky.social in which we built a platform-agnostic framework that models and connects users as distributions over latent narratives rather than interactions.

1

Patrick Gerard @patrikgerard.bsky.social · Jun 22

In my talk, I'll discuss my current work with @emilioferrara.bsky.social aimed at understanding the cross-platform toxic narrative movement online.

1 1

Patrick Gerard @patrikgerard.bsky.social · Jun 22

🧵 FINALLY, excited to have been invited to present a pitch talk at the #DETOX workshop at @icwsm.bsky.social 2025, organized by @alessianetwork.bsky.social , @lajello.bsky.social l, Alessandro Galeazzi, & Eugenia Polizzi. Link to the workshop is here! detox-workshop.github.io/website.gith...

DETOX@ICWSM2025

detox-workshop.github.io

1 1 4

Patrick Gerard @patrikgerard.bsky.social · Jun 22

Paper here:
arxiv.org/abs/2505.21729

Bridging the Narrative Divide: Cross-Platform Discourse Networks in Fragmented Ecosystems

Political discourse has grown increasingly fragmented across different social platforms, making it challenging to trace how narratives spread and evolve within such a fragmented information ecosystem....

arxiv.org

2 6

Patrick Gerard @patrikgerard.bsky.social · Jun 22

This is a new way of thinking about user networks: birds of the same feather don't simply tweet together; they *think* together. Connecting users via latent narratives allows us to capture more users with less data and break out of arbitrary platform boundaries.

1

Patrick Gerard @patrikgerard.bsky.social · Jun 22

When we apply this method to X/Truth Social during the 2024 US Presidential Election, we find a mixed-platform group comprising just 0.33% of users who were associated with ~70% of narrative migration between platforms: a subgroup effectively invisible to other network construction methods.

1

Patrick Gerard @patrikgerard.bsky.social · Jun 22

⬆️ This represents a modernization of user modeling and user network construction in times marked by increasingly restricted API access and deeply fragmented information ecosystems.

1

Patrick Gerard @patrikgerard.bsky.social · Jun 22

⬆️ This lets us create platform-agnostic user networks that capture way more users than traditional methods, achieve competitive performance on classic CSS network tasks with significantly less data, and substantially outperform current approaches for cross-platform info diffusion modeling.

1

Patrick Gerard @patrikgerard.bsky.social · Jun 22

While I'm here, I might as well post about my newest work with @luceriluc.bsky.social and @emilioferrara.bsky.social in which we built a platform-agnostic framework that models and connects users as distributions over latent narratives rather than interactions.

1 3

Patrick Gerard @patrikgerard.bsky.social · Jun 22

Real-world Impact: We can now examine the mechanisms behind "us vs them" construction in real-time: how communities systematically build threat narratives through in-group language.

Patrick Gerard @patrikgerard.bsky.social · Jun 22

Interesting Findings From Our Analysis:

- Othering messages get significantly more attention/engagement.

- Othering surges during crises, and gets even more attention.

- Othering shows intersection with moral language: "moralized othering" is effective propaganda.

1

Patrick Gerard @patrikgerard.bsky.social · Jun 22

Cross-Domain Transfer: We introduce "Rapid Domain Adaptation", finding that we can fine-tune on one domain and adapt to completely different contexts with just system prompt steering + logit disambiguation, achieving 90%+ accuracy across languages and cultures without retraining.

1

Patrick Gerard @patrikgerard.bsky.social · Jun 22

Cheaper Training: Exploiting decoder-based capabilities, we introduce a scalable "Artificial Annotator Alignment" pipeline, which needs only a fraction of the human annotations traditional approaches need.

1

Patrick Gerard @patrikgerard.bsky.social · Jun 22

Novel Modeling Capabilities: We develop 4 categories grounded in sociological theory that capture different stages of the othering process which underlies hate speech. This helps move beyond detecting expressions of hate to understanding how targets are systematically constructed as threats.

1

Patrick Gerard @patrikgerard.bsky.social · Jun 22

Methodological Progress ⬇️

1

Patrick Gerard @patrikgerard.bsky.social · Jun 22

For example: "Banderites isn't in any hate speech dictionary, but LLMs understand it's a historically-loaded term used to create the perception of threat.

1

Patrick Gerard @patrikgerard.bsky.social · Jun 22

Major Insight: Due to its complexity and nuance, Traditional decoder models perform terribly on detecting othering language. But LLMs excel at understanding in-group speak - the coded, context-dependent language communities use.

1

Patrick Gerard @patrikgerard.bsky.social · Jun 22

Why Care? Traditional (hate|fear) speech detection misses the nuanced, context-dependent language that historically justified violence -- from Nazi Germany to Myanmar's Rohingya Genocide. Othering is the broader psychological process that hate speech serves.

1

Patrick Gerard @patrikgerard.bsky.social · Jun 22

🧵 ALSO Excited to present our work on "othering" at @icwsm.bsky.social 2025! In this paper with @kristinalerman.bsky.social, we go beyond traditional (hate|fear) speech to identify and study the mechanisms of "othering" language: the subtle process of depicting outgroups as fundamentally different.

1 2 2

Patrick Gerard @patrikgerard.bsky.social · Jun 22

Why Care? Traditional approaches provide snapshots; our method captures the evolutionary process itself. 93.2% of trending clusters correspond to documented external events, validating our temporal detection capabilities!

Patrick Gerard @patrikgerard.bsky.social · Jun 22

Cool (?) Example: The "biological weapons labs" narrative demonstrates traceable evolution dynamics. After Russian UN accusations, we observe branching into sub-narratives about "new documents," "US-financed experiments," and "bat coronavirus samples."

1

Patrick Gerard @patrikgerard.bsky.social · Jun 22

Interesting finding:s Analyzing 9.67M Telegram posts, Russian communities spike on US aid announcements (framing as Western aggression), while Ukrainians react to battlefield wins/losses. Same war, parallel evolutionary trees of information.

1

Patrick Gerard @patrikgerard.bsky.social · Jun 22

Key insight: Narratives exhibit evolutionary dynamics - they emerge, adapt to new information, split into sub-narratives, and merge during critical events. Traditional static methods miss these temporal dynamics entirely.

1

Patrick Gerard @patrikgerard.bsky.social · Jun 22

Core contribution: A new method for finding+tracking narratives on-the-fly that treats information narratives as dynamic, evolving structures rather than static topic clusters. Our method adapts hierarchical clustering for real-time data streams, tracking narratives across time.

1

Patrick Gerard @patrikgerard.bsky.social · Jun 22

🧵 Excited to present at @icwsm.bsky.social 2025! In this paper with @kristinalerman.bsky.social, we introduce a novel framework for modeling narrative evolution in streaming text data. Link here: arxiv.org/abs/2409.07684

1 4 5