Language Predicts Identity Fusion Across Cultures and Reveals Divergent Pathways to Violence

In light of increasing polarization and political violence, understanding the psychological roots of extremism is increasingly important. Prior research shows that identity fusion predicts willingness to engage in extreme acts. We evaluate the Cognitive Linguistic Identity Fusion Score, a method that uses cognitive linguistic patterns, LLMs, and implicit metaphor to measure fusion from language. Across datasets from the United Kingdom and Singapore, this approach outperforms existing methods in predicting validated fusion scores. Applied to extremist manifestos, two distinct high-fusion pathways to violence emerge: ideologues tend to frame themselves in terms of group, forming kinship bonds; whereas grievance-driven individuals frame the group in terms of their personal identity. These results refine theories of identity fusion and provide a scalable tool aiding fusion research and extremism detection.

💡 Research Summary

The paper tackles the pressing problem of detecting psychological precursors to political violence by focusing on identity fusion, a strong alignment between an individual’s self and a target group, belief, or ideology that has been linked to willingness to engage in extreme acts. The authors introduce the Cognitive Linguistic Identity Fusion Score (CLIFS), a novel measurement that combines four complementary components: (1) a set of cognitively informed linguistic patterns identified in prior fusion research; (2) opaque embedding features derived from a fine‑tuned ModernBERT model to capture subtle semantic signals beyond the reach of hand‑crafted rules; (3) coarse‑grained fusion‑level probabilities from a fine‑tuned classifier that distinguishes “high‑fusion” from “low‑fusion” text; and (4) implicit metaphor detection using a masked language model, which yields three key metrics—Fictive‑Kinship (Kf), Fusion Proximity f(I,T), and directional identity scores (SI→T and ST→I). These metaphor‑based features have previously shown the greatest performance gains in CLIFS ablation studies, outperforming high‑dimensional semantic embeddings.

To evaluate CLIFS, the authors use three out‑of‑domain datasets that span cultural contexts (the United Kingdom and Singapore), collection methods (Amazon Mechanical Turk surveys and in‑person church interviews), and domains (religious conversion narratives and extremist manifestos). The first dataset contains 85 participants with Verbal Identity Fusion Scale (VIFS) scores; the second comprises 33 interview transcripts; the third consists of 18 violent manifestos (nine previously studied and nine newly collected, split into “Ideologue” and “Victim” categories). Because CLIFS requires a minimum amount of textual context, the authors filter short responses using word‑ and sentence‑count thresholds (30–50 words, 2–4 sentences) and chunk longer texts into 300‑word, sentence‑preserving segments to stay within the 384‑token limit of the underlying sentence transformer.

Baseline comparators include the Unquestioning Affiliation Index (UAI) and its simplified version (nUAI), both of which rely on linguistic inquiry and word‑count features, and the Violence Risk Index Fusion (VRI‑Fusion), a dictionary‑based metric that counts kinship words. Across all filtering conditions, CLIFS consistently achieves the highest Spearman correlations with ground‑truth VIFS scores (rₛ = 0.31–0.55, all p < 0.05) and the lowest mean absolute error (MAE = 1.11–1.75). In contrast, UAI, nUAI, and VRI‑Fusion either fail to reach statistical significance or produce substantially larger errors (MAE ranging from ~3.9 to >50). The correlation strength improves as stricter length thresholds are applied, indicating that richer contextual information benefits the model.

The manifesto analysis explores two hypothesized pathways to violence: an “Ideologue” pathway, where individuals fuse with a broader ideological group, and a “Victim” pathway, where personal grievance drives violent action without a clear group identity. While overall CLIFS scores do not differ dramatically between the two groups (mean difference ≈ 0.05, Cohen’s d ≈ 0.06), the implicit‑metaphor components reveal distinct patterns. Victim‑driven texts show higher Fusion Proximity and higher ST→I scores, suggesting a projection of the self onto the target (“I see the group as like me”). Ideologue texts exhibit higher Fictive‑Kinship values, indicating a familial framing of the group (“we are family”). Directional SI→T scores are modestly higher for ideologues, reflecting a stronger self‑to‑group identification. These findings support a nuanced view of identity fusion: it can be asymmetric and not always kin‑based, leading to multiple routes toward violent behavior.

The discussion situates the results within the Comprehensive Identity Fusion Theory (CIFT), which has recently expanded the concept of fusion beyond concrete social groups to include abstract values and ideologies. CLIFS proves sensitive to these broader constellations, reinforcing the theoretical shift. Practically, CLIFS offers a scalable, language‑agnostic tool for early detection of high‑fusion individuals in massive online streams, overcoming the limitations of prior rule‑based or keyword‑matching approaches that suffer from overfitting, high false‑positive rates, and poor cross‑cultural transferability. The authors also critique the VRI‑Fusion method, highlighting its reliance on manual dictionary construction and lack of reproducibility, which the present study confirms as ineffective on out‑of‑domain data.

In sum, the paper demonstrates that (1) CLIFS accurately predicts validated identity‑fusion scores across culturally diverse, out‑of‑domain text; (2) it outperforms existing linguistic fusion metrics; and (3) it uncovers two distinct metaphor‑driven pathways to extremist violence. The work paves the way for deploying CLIFS in real‑time monitoring systems, extending research to non‑English corpora, and informing policy interventions aimed at mitigating radicalization before it translates into violent action.