Introduction
Identity fusion represents a psychological state where the boundaries between an individual's personal identity and a group identity become porous, creating a "visceral feeling of oneness." Unlike social identification, which involves a sense of belonging, fusion drives extreme pro-group behaviors, including self-sacrifice. Traditionally, measuring fusion required direct psychological surveys, which are difficult to scale for large-scale digital analysis. This paper introduces the Cognitive Linguistic Identity Fusion Score (CLIFS), a novel, scalable approach to quantifying identity fusion from raw text.
Methodology: The CLIFS Framework
The core innovation of CLIFS lies in its use of Implicit Metaphor Detection via Masked Language Modeling (MLM). The researchers hypothesize that fused individuals subconsciously treat "self" and "group" concepts as interchangeable in their writing.
As shown in the architecture diagram above, the pipeline begins with data augmentation using Round-Trip Translation (RTT) and Generative AI to address the scarcity of labeled psychological data. The model then extracts several key features:
- Directional Proximity ($S_{I \to T}$ and $S_{T \to I}$): Using ModernBERT, the model masks "self" tokens (e.g., "I", "me") and calculates the probability of predicting "group" tokens (e.g., "we", "[Target]") in their place, and vice versa.
- Bidirectional Proximity ($f_{I,T}$): A harmonic mean of the directional scores captures the reciprocity of the bond—how much the personal self and the group identity reinforce each other.
- Fictive Kinship ($K_f$): This metric measures the linguistic framing of the group as a family unit (e.g., "brothers in arms").
- Ensemble Classification: These scores are combined with SBERT embeddings and traditional lexical markers (such as the Unquestioning Affiliation Index) into a Random Forest classifier to produce the final fusion label.
Empirical Validation
The effectiveness of these metrics is demonstrated through their distribution across different levels of fusion. The researchers validated the model using an Identity Fusion Dataset of 871 essays and a Violence Risk Benchmark consisting of ideological manifestos.
As illustrated in the CDF plots, the derived metrics show clear separation between "low," "medium," and "high" fusion labels. The bidirectional proximity score ($f_{I,T}$) and fictive kinship ($K_f$) are particularly effective at distinguishing highly fused individuals from those with moderate or low identification, confirming that these linguistic signals correlate strongly with psychological states.
Key Results and Performance
CLIFS demonstrates significant improvements over existing automated and manual methods:
- Superior Accuracy: CLIFS outperformed existing automated baselines by 6–154% in classification tasks.
- Human-Level Performance: Remarkably, the model surpassed human annotators by 11–22%. This suggests that the MLM-based approach identifies subtle linguistic patterns—subconscious metaphorical substitutions—that human readers often overlook.
- Violence Prediction Powerhouse: When integrated into a violence risk assessment framework, CLIFS provided a 240% predictive gain. It successfully identified the "Violent Self-Sacrificial" nature of manifestos without the extensive manual filtering required by previous tools.
Discussion and Ethical Considerations
While CLIFS offers a powerful tool for computational social science, the authors highlight critical ethical risks. Such technology could potentially be weaponized by authoritarian regimes for surveillance or marginalization of specific groups. Consequently, the researchers advocate for its use as an augmentative tool for human experts in high-stakes environments, rather than a standalone decision-maker.
Future work will focus on expanding the model's cultural and linguistic diversity, as the current iteration is primarily focused on English-language, U.S.-centric data. The team also plans to explore longitudinal tracking to understand how fusion scores change over time during processes of radicalization or defusion.
Conclusion
CLIFS represents a significant step forward in bridging the gap between deep psychological theory and large-scale natural language processing. By moving beyond keyword matching to capture the contextual interchangeability of self and group, CLIFS provides a robust, cognition-informed metric for understanding the linguistic signatures of identity fusion.
Summary updated