TL;DR
The paper introduces the Cognitive Linguistic Identity Fusion Score (CLIFS), a novel, fully automated metric that uses masked language models to quantify "identity fusion"—the psychological merging of oneself with a group or ideology—directly from text. By capturing implicit metaphors in natural language, CLIFS eliminates the need for intrusive surveys, outperforms human annotators, and improves downstream violence risk assessment by over 240%.
What Problem Does This Paper Address?
Identity fusion is a psychological state where an individual's personal identity becomes deeply intertwined with a target identity (like a religion, political party, or ideology), often leading to extreme behaviors ranging from deep prosociality to violent self-sacrifice.
Traditionally, psychologists measure identity fusion using pictorial scales (overlapping circles) or verbal questionnaires. While effective, these methods require direct, controlled surveys, making it impossible to analyze large-scale, historical, or real-time text data (e.g., online forums or manifestos). Existing automated text-based metrics rely heavily on explicit dictionary matching, which correlates poorly with gold-standard psychological scales. The authors address this gap by creating a scalable, text-only metric that accurately reflects an individual's subconscious level of identity fusion.
Methods
CLIFS operates on the premise that highly fused individuals subconsciously use identity tokens (e.g., "I", "me") and target tokens (e.g., "religion", "country") interchangeably.
Instead of looking for explicit keywords, CLIFS leverages a masked language model (ModernBERT) to detect implicit metaphors through token prediction. If identity tokens are masked out, how likely is the model to suggest the fusion target in their place?
The architecture incorporates several key components:
- Bidirectional Proximity: Because identity fusion is reciprocal ("I am one with my group" and "My group is me"), CLIFS calculates directional proximity from identity to target ($S_{I\rightarrowT}$) and vice versa ($S_{T\rightarrowI}$), combining them via a harmonic mean ($f_{(I,T)}$).
- Fictive Kinship ($K_f$): Fused individuals often view their group as family. CLIFS measures this by calculating the proximity between kinship terms (e.g., "brother", "family") and the fusion target.
- Machine Learning Pipeline: These core linguistic metrics are combined with SBERT embeddings, lexical markers from prior indices, and fine-tuned ModernBERT class probabilities into a Random Forest classifier and regressor.
To combat data scarcity, the authors utilized two forms of AI data augmentation: Round-Trip Translation (RTT) via German and Chinese, and synthetic text generation using GPT-4o.
Datasets
The model was trained and evaluated using two distinct datasets:
- Identity Fusion Dataset: Sourced from prior work by Ashokkumar and Pennebaker (2022), containing essays from 871 MTurk participants writing about their relationship to a country, religion, or university. Ground truth was established using the gold-standard Verbal Identity Fusion Scale (VIFS).
- Violence Risk Prediction Dataset: A corpus of 15 ideological manifestos (including texts from mass shooters and historical figures like Adolf Hitler and Martin Luther King Jr.) segmented into 1,971 balanced, 300-word chunks, categorized as "Violent Self-Sacrificial," "Ideologically Extreme," or "Moderate."
Key Insights and Ideas
- Subconscious Interchangeability: The core "aha moment" is that masked language models can capture the psychological proximity of concepts. Fused individuals frame their speech such that "self" and "target" occupy the same conceptual space, a nuance that dictionary-matching methods completely miss.
- Reciprocity is Essential: Modeling identity fusion as a bidirectional relationship significantly improves alignment with human psychological reality compared to unidirectional language tracking.
- Linguistic Shifts: Empirical analysis of the cumulative distribution functions (CDFs) confirmed that as true fusion levels increase, there is a distinct distributional shift in how interchangeably identity and target words are predicted by the LLM.
Key Results
- Outperforming Humans: In fine-grained identity fusion classification (Experiment 1), the CLIFS Random Forest and Ensemble models achieved a macro $F_1$ score of 0.66, outperforming baseline automated methods by 6–154% and beating human text annotation by 11–22%.
- Strong Correlation: The CLIFS regression model achieved a Spearman correlation of $r_s = 0.633$ ($p \ll 0.001$) against actual VIFS scores, representing a massive predictive gain over previous linguistic indices.
- Downstream Impact: When integrated into a fully automated violence risk assessment pipeline, replacing the legacy fusion module with CLIFS features boosted the macro $F_1$ score from 0.18 to 0.62—a 240%+ predictive gain. Notably, Fictive Kinship ($K_f$) emerged as the single strongest individual predictor of violence risk.
Limitations and Future Work
While powerful, CLIFS currently has a few limitations. The non-synthetic training data is relatively small and restricted to three fusion targets (country, religion, university). Additionally, the dataset relies entirely on English-language texts from U.S. participants, meaning the model's features may not seamlessly transfer to other cultural contexts without multilingual validation.
The authors also note ethical risks: if misinterpreted or weaponized by authoritarian regimes, automated fusion scoring could be used to unfairly target or marginalize vulnerable groups, mistaking high group cohesion for imminent violence. Future work will focus on expanding to cross-cultural corpora and exploring feature interactions using advanced interpretability frameworks like SHAP.
Summary updated