Cognitive Linguistic Identity Fusion Score (CLIFS)

• Cognitive Linguistic Identity Fusion Score (CLIFS) is an automated metric that computes the alignment between self and target in text using cognitive-linguistic techniques.
• It leverages masked language models and implicit metaphor detection to capture reciprocal self–target proximity, offering both coarse-grained classification and fine-grained regression outputs.
• CLIFS enables scalable, large-scale analysis compared to traditional survey methods, demonstrating improved reproducibility and cross-cultural validation in measuring identity fusion.

Searching arXiv for the specified CLIFS papers to ground the article in current preprints. Cognitive Linguistic Identity Fusion Score (CLIFS) is an automated, text-based metric for estimating identity fusion from natural language. Identity fusion is framed as the “visceral feeling of oneness” between the self and a target such as a group, country, religion, university, ideology, leader, value, brand, belief, or other abstract target. CLIFS was introduced as a scalable, fully automated alternative to survey-based and manually coded measures, integrating cognitive linguistics with LLMs and building on implicit metaphor detection in order to infer how a speaker conceptually aligns self and target in text (Wright et al., 20 Sep 2025, Wright et al., 9 Feb 2026).

1. Identity fusion as the measured construct

Identity fusion denotes a psychological merging of personal identity with another entity or abstract target. In the CLIFS framing, the defining issue is not merely whether a speaker expresses approval of a target, but whether the language reveals that self and target are emotionally and conceptually intertwined. The construct matters because prior research on fusion theory links high fusion to willingness to make extreme sacrifices, including violence, on behalf of the fused target (Wright et al., 9 Feb 2026).

CLIFS adopts a specifically cognitive-linguistic view of the construct. Its central assumption is that, in high identity fusion, self and target become conceptually close, so words referring to one should be predictively interchangeable with words referring to the other in context. This shifts measurement away from explicit self-report and toward implicit conceptual framing. The relevant signal is therefore reciprocal self–target proximity, often expressed through metaphorical identity blending, kinship framing, and self–group interchangeability rather than only through overt declarations of loyalty (Wright et al., 20 Sep 2025).

The broader theoretical background, as described in the later cross-cultural study, is compatible with the Comprehensive Identity Fusion Theory view that fusion can extend beyond ordinary groups to values, ideologies, leaders, and abstract targets. This suggests that CLIFS is intended not only for conventional group attachment but also for cases in which the fused object is doctrinal, symbolic, or otherwise non-corporeal (Wright et al., 9 Feb 2026).

2. Position within the measurement landscape

CLIFS was proposed in response to limitations in existing identity-fusion measurement. Traditional pictorial measures, including DIFI, rely on circle-overlap judgments, are survey-based, and require direct participant interaction. VIFS, the Verbal Identity Fusion Scale, is the gold standard and is computed as the mean of seven item ratings, but it still requires explicit responses collected from participants. Automated alternatives were presented as inadequate for large-scale, ecologically valid deployment: UAI showed only weak-to-moderate alignment with VIFS, while VRI was characterized as measuring violence-risk linguistics more than direct fusion and as not aligning with VIFS (Wright et al., 20 Sep 2025).

CLIFS differs by operating directly on free text. It is fully automated, uses cognitive linguistics plus masked LLMs, is built around implicit metaphor detection rather than explicit questionnaire items, and models fusion as a reciprocal relation between self and target language. The later paper further positions it against approaches that rely on simple dictionary matching, manually curated keywords, assumptions that do not transfer across data distributions, and manual post-processing that undermines scalability and reproducibility (Wright et al., 9 Feb 2026).

Measure	Basis	Limitation stated in the sources
Pictorial Measure / DIFI	Circle-overlap judgments	Survey administration and direct participant interaction
VIFS	7-item verbal scale; mean of item ratings	Requires explicit human response collection
UAI	Automated baseline	Weak-to-moderate alignment with VIFS
VRI / VRI-Fusion	Violence-risk linguistics / fusion component	Does not align with VIFS
Human annotation	Manual coding	Expensive, small-scale, subject to annotator subjectivity

Because CLIFS is intended to estimate the VIFS score an individual would have received from written text alone, it functions as a text-only proxy for psychometric identity fusion. The papers describe two operational modes: coarse-grained classification into low, medium, or high fusion, and fine-grained regression to a continuous score on the 1–7 VIFS scale (Wright et al., 9 Feb 2026).

3. Cognitive-linguistic architecture and feature construction

The CLIFS pipeline begins with natural-language text written by a person about a fusion target. In the original benchmark, examples included essays about country, religion, and university; in downstream violence-risk work, the inputs were sentence-preserving manifesto segments of approximately 300 words (Wright et al., 20 Sep 2025). In the later cross-cultural work, longer texts were also chunked because the Sentence Transformer setup based on all-mpnet-base-v2 had a 384-token input limit (Wright et al., 9 Feb 2026).

Three lexical sets anchor the method. The identity vocabulary $I$ comprises first-person singular pronouns such as I, me, my, mine, myself. The fusion target vocabulary $T$ contains first-person plural pronouns and target terms, including items such as we, us, our, and is parameterized to the current target domain. The kinship vocabulary $K$ contains family and kin terms such as brother, sister, family, kin, brethren, comrades. These seed lists are expanded with GloVe embeddings by adding all vocabulary items with cosine similarity greater than $0.8$ to the seed terms, and spaCy NER is used so that organizations, nationalities/religious/political groups, and geopolitical entities can be masked as target-like references when computing certain scores (Wright et al., 20 Sep 2025).

The distinctive mechanism is masked-LM implicit metaphor detection. CLIFS masks one category of words in a document and uses ModernBERT to estimate what words are likely in that context. If masking target words produces high probability mass on identity words or kinship words, the language is taken to indicate conceptual proximity between self and target. The method is explicitly described as inspired by prior masked-LM metaphor detection work, including work on immigration speeches and dehumanizing language (Wright et al., 20 Sep 2025, Wright et al., 9 Feb 2026).

Directional proximity is formalized as

$$S_{x \to y} = \frac{1}{M_y} \sum_{m=1}^{M_y} \sum_{w_v \in \mathcal{V}_x} P\bigl(w_v \mid C_m\bigr)^\alpha$$

where $x, y \in \{I, T\}$, $M_y$ is the number of masked positions for category $y$, $C_m$ is the masked context at position $m$, $T$0 is the vocabulary for category $T$1, and $T$2 is the masked LLM probability of candidate word $T$3. The smoothing exponent satisfies $T$4, and the paper uses $T$5. The stated reason for the exponent is that raw LM probabilities are tiny, so raising them to $T$6 increases separability and stabilizes aggregation (Wright et al., 20 Sep 2025).

Reciprocity is central to the model. CLIFS computes $T$7, in which identity terms are predicted when target terms are masked, and $T$8, in which target terms are predicted when identity terms are masked. These are combined using a harmonic-mean-style formula, analogous to $T$9, to form fusion proximity $K$0. The method therefore penalizes asymmetric cases in which only one direction is strong. A separate fictive kinship score is defined as

$K$1

capturing the degree to which kinship terms are evoked when target words are masked or considered (Wright et al., 20 Sep 2025, Wright et al., 9 Feb 2026).

In the full predictive pipeline, CLIFS is not treated as a single raw scalar but as a feature suite. The sources list $K$2, $K$3, $K$4, $K$5, and classifier output probabilities or predicted class as CLIFS-derived features. These are combined with lexical baselines such as UAI and VRI outputs, SBERT embeddings, and a fine-tuned ModernBERT classifier’s soft class probabilities, then supplied to classical machine-learning models including random forest, SVM, XGBoost, and an ensemble (Wright et al., 20 Sep 2025).

4. Initial benchmarks, augmentation, and comparison with human coding

The original benchmark used an identity-fusion prediction dataset derived from Ashokkumar and Pennebaker. It comprised 871 MTurk participants who wrote 6–8 minute essays about their relationship to country ($K$6), religion ($K$7), or university ($K$8), with VIFS scores as ground truth. For classification, the VIFS scores were discretized into low, medium, and high using standard deviation cutoffs from the mean. An important detail is that the country condition contained only 4 of the 7 VIFS items (Wright et al., 20 Sep 2025).

To compensate for the small identity-fusion dataset, the study used round-trip translation augmentation through English $K$9 German $0.8$0 English and English $0.8$1 Chinese $0.8$2 English, as well as GPT-4o-generated synthetic fusion texts and some minority-class oversampling. The paper reports that augmentation improved trainable models by about 10% on average (Wright et al., 20 Sep 2025).

The benchmark compared CLIFS with majority vote, zero-shot ModernBERT, few-shot GPT-4o, GPT-4o RAG, DeepSeek R1 RAG, SBERT plus random forest, fine-tuned ModernBERT, XGBoost, SVM, UAI, VRI, and a single human annotator. On augmented data, the best overall macro $0.8$3 in identity-fusion classification was $0.8$4 for both CLIFS Ensemble and CLIFS RF. Bootstrapped confidence intervals were reported as $0.8$5 with 95% CI $0.8$6 for CLIFS RF and $0.8$7 with 95% CI $0.8$8 for CLIFS Ensemble. Comparator values included majority vote at $0.8$9, zero-shot at $$S_{x \to y} = \frac{1}{M_y} \sum_{m=1}^{M_y} \sum_{w_v \in \mathcal{V}_x} P\bigl(w_v \mid C_m\bigr)^\alpha$$0, SBERT RF at $$S_{x \to y} = \frac{1}{M_y} \sum_{m=1}^{M_y} \sum_{w_v \in \mathcal{V}_x} P\bigl(w_v \mid C_m\bigr)^\alpha$$1–$$S_{x \to y} = \frac{1}{M_y} \sum_{m=1}^{M_y} \sum_{w_v \in \mathcal{V}_x} P\bigl(w_v \mid C_m\bigr)^\alpha$$2 depending on condition, ModernBERT up to $$S_{x \to y} = \frac{1}{M_y} \sum_{m=1}^{M_y} \sum_{w_v \in \mathcal{V}_x} P\bigl(w_v \mid C_m\bigr)^\alpha$$3, and GPT-4o RAG around $$S_{x \to y} = \frac{1}{M_y} \sum_{m=1}^{M_y} \sum_{w_v \in \mathcal{V}_x} P\bigl(w_v \mid C_m\bigr)^\alpha$$4. The paper summarizes these results as a 6–154% improvement over baselines (Wright et al., 20 Sep 2025).

Regression results showed a best augmented-data performance of MAE $$S_{x \to y} = \frac{1}{M_y} \sum_{m=1}^{M_y} \sum_{w_v \in \mathcal{V}_x} P\bigl(w_v \mid C_m\bigr)^\alpha$$5 with Spearman $$S_{x \to y} = \frac{1}{M_y} \sum_{m=1}^{M_y} \sum_{w_v \in \mathcal{V}_x} P\bigl(w_v \mid C_m\bigr)^\alpha$$6. The reported baseline correlations were $$S_{x \to y} = \frac{1}{M_y} \sum_{m=1}^{M_y} \sum_{w_v \in \mathcal{V}_x} P\bigl(w_v \mid C_m\bigr)^\alpha$$7 for UAI and $$S_{x \to y} = \frac{1}{M_y} \sum_{m=1}^{M_y} \sum_{w_v \in \mathcal{V}_x} P\bigl(w_v \mid C_m\bigr)^\alpha$$8 for VRI-fusion, leading the paper to report a 165–419% gain in correlation strength in the test setup and gains up to 1.3–29× on the full dataset (Wright et al., 20 Sep 2025).

Human comparison was conducted on 97 college-target samples. The source explicitly notes that this benchmark used only one annotator, who was college-educated and likely informed about the construct, so it is not a multi-rater gold standard. Reported macro $$S_{x \to y} = \frac{1}{M_y} \sum_{m=1}^{M_y} \sum_{w_v \in \mathcal{V}_x} P\bigl(w_v \mid C_m\bigr)^\alpha$$9 was $x, y \in \{I, T\}$0 for the annotator, $x, y \in \{I, T\}$1 for CLIFS RF, and $x, y \in \{I, T\}$2 for the CLIFS ensemble, which the paper characterizes as an 11–22% gain over human annotation. In regression on the same comparison set, CLIFS achieved MAE $x, y \in \{I, T\}$3 and $x, y \in \{I, T\}$4, whereas human ratings versus VIFS had MAE $x, y \in \{I, T\}$5 and $x, y \in \{I, T\}$6, corresponding to an error reduction of about 25% in that comparison (Wright et al., 20 Sep 2025).

Ablation results indicated that all feature groups mattered. The paper states that CLIFS identity-fusion features produced the largest gain among non-trained features, ModernBERT class probabilities provided a strong boost, SBERT features dominated splits in the random forest, and interpretable CLIFS features still contributed importantly. This suggests that the method is hybrid in a substantive sense: cognitively motivated variables and opaque representation learning were both empirically consequential (Wright et al., 20 Sep 2025).

5. Cross-cultural validation and divergent pathways to violence

The second CLIFS paper focused on out-of-domain and cross-cultural validation. One dataset came from prior unpublished work on identity fusion and significant religious experience in the United Kingdom and Singapore. After exclusions for incomplete surveys and lack of significant religious experience, 521 participants remained; the CLIFS analysis used the subset who wrote about conversion or reaffirmation experiences and also had VIFS scores, yielding $x, y \in \{I, T\}$7. Because many responses were short, the authors tested multiple filtering conditions: 30, 40, and 50-word thresholds, which produced $x, y \in \{I, T\}$8, and 2, 3, and 4-sentence thresholds, which produced $x, y \in \{I, T\}$9 (Wright et al., 9 Feb 2026).

A second dataset came from eight months of fieldwork in Singapore at two churches. Participants completed semi-structured interviews in English about Sunday services. After exclusions, the final sample was $M_y$0. For CLIFS evaluation, the study analyzed both the full transcript as one item and a chunked version generated to fit the embedding model’s input limit, producing $M_y$1 chunks, later filtered to $M_y$2 with at least four sentences (Wright et al., 9 Feb 2026).

Against true VIFS scores, CLIFS outperformed UAI, nUAI, and VRI-Fusion in these out-of-domain settings. In the MTurk religious-experience dataset, CLIFS reached $M_y$3 to $M_y$4, depending on filtering, with all correlations significant at $M_y$5, while the baselines did not achieve significant correlations in any MTurk condition. In the Singapore field dataset, CLIFS again outperformed the baselines, with $M_y$6, $M_y$7 for full transcripts and $M_y$8, $M_y$9 for chunked analysis. UAI reached significance only in one full-transcript condition at $y$0, $y$1, and did not generalize to chunked data; nUAI and VRI-Fusion were not significant. MAE values also favored CLIFS strongly: CLIFS ranged from $y$2 to $y$3, UAI from $y$4 to $y$5, VRI-Fusion from $y$6 to $y$7, and nUAI from $y$8 to $y$9 (Wright et al., 9 Feb 2026).

The same paper applied CLIFS to an extremist manifesto corpus expanded from prior work. It retained nine manifestos previously labeled “Violent Self-Sacrificial” or “Ideologically Extreme” and added nine more violent manifestos collected by CulturePulse, yielding 18 manifestos with labels collapsed into Ideologue and Victim. The paper lists examples including Payton Gendron, Theodore Kaczynski, Seung-hui Cho, Christopher Dorner, Eric Harris and Dylan Klebold, Elliot Rodger, Brenton Tarrant, Dylann Roof, Anders Behring Breivik, Adolf Hitler, and Sayyid Qutb (Wright et al., 9 Feb 2026).

The manifesto analysis did not strongly support the expectation that ideologues would show stronger overall identity fusion than grievance-driven actors. Instead, CLIFS score distributions were broadly similar, with only small differences and slightly higher fusion in the Victim class. The more consequential result emerged from inspection of high-fusion chunks. Ideological manifestos showed stronger fictive kinship, indicating that the group was framed as family and that the writer’s identity was linguistically aligned with collective membership. Victimhood-driven texts showed higher fusion proximity $C_m$0 and higher $C_m$1, suggesting that the group or target was framed in terms of personal identity rather than principally through kin-based solidarity. The authors interpret these as two distinct high-fusion pathways to violence: a kinship-based, group-centered route and a grievance-driven self-projection route (Wright et al., 9 Feb 2026).

These findings refine the theory in two ways stated in the paper. First, fusion is not expressed through a single linguistic pattern. Second, threat need not be overtly existential; social rejection, symbolic humiliation, or conceptual “death” of an idealized self may suffice to activate violence among fused individuals. A plausible implication is that CLIFS is useful not only because it predicts validated scores, but because it decomposes fusion into directional and kinship-sensitive components that can separate distinct violence-relevant psycholinguistic profiles (Wright et al., 9 Feb 2026).

6. Applications, constraints, and prospective development

The original CLIFS paper presented violence-risk assessment as a proof of concept. Using the VRI framework, it replaced the original fusion component with CLIFS-derived features: $C_m$2, $C_m$3, $C_m$4, $C_m$5, and a CLIFS RF prediction, then trained a random forest for violence-risk classification. Reported macro $C_m$6 values were $C_m$7 for majority vote, $C_m$8 for the original VRI, $C_m$9 for a random forest on VRI submodules, and $m$0 for VRI + CLIFS. The paper explains the “more than 240%” improvement by comparing $m$1 with $m$2, which yields approximately $m$3, or a 244% relative improvement. It also stresses that this benchmark omitted the manual filtering step used in the original VRI paper, making the comparison more realistic as an automated benchmark (Wright et al., 20 Sep 2025).

The papers assign broader significance to CLIFS because it enables large-scale identity-fusion measurement, longitudinal tracking, historical corpus analysis, counter-terrorism and extremism research, violence-risk evaluation, cultural analytics, and the study of prosocial as well as risky group behavior. The original study also suggests possible use in monitoring shifts in group cohesion or defusion over time (Wright et al., 20 Sep 2025). The later paper adds that CLIFS can aid large-scale analysis of interviews, manifestos, and online text and may be relevant for triage or early-warning systems for high-risk content (Wright et al., 9 Feb 2026).

Several limitations are explicit. The human benchmark in the original paper rests on a single annotator, so no inter-rater reliability is available. The training data for identity-fusion learning were small and narrow: only 873 non-synthetic entries and only three fusion targets—country, religion, and university. The authors caution that performance may not transfer well to social movements, brands, online communities, other target types, or other contexts. The data are English and U.S.-centric in the first paper, and the authors call for multilingual and cross-cultural corpora. Feature-importance analysis used Gini Importance, which the paper notes does not capture interactions well; SHAP and TreeSHAP are suggested for future work (Wright et al., 20 Sep 2025).

Ethical caveats are similarly explicit. CLIFS should not replace human judgment in high-stakes settings. Misclassification could be harmful, and high fusion does not necessarily imply violence because it can also correlate with prosocial or cooperative behavior. The original paper warns about possible misuse by authoritarian or malicious actors (Wright et al., 20 Sep 2025). The later paper also notes that some evaluation datasets are small, chunking may alter context, and the extremist-manifesto labels are coarse, reducing a complex motivational landscape to two classes (Wright et al., 9 Feb 2026).

Future work is described in concrete terms: extension to more fusion-target domains, larger and more diverse datasets, multilingual settings, cross-cultural samples, improved feature attribution methods, and out-of-domain validation (Wright et al., 20 Sep 2025). The original study further suggests that a fast local random-forest implementation may be preferable in most use cases, while the ensemble may be more useful for out-of-domain settings if local DeepSeek hosting is available (Wright et al., 20 Sep 2025). Taken together, the two papers present CLIFS as a cognitively informed language-based measure that estimates VIFS from text alone, scales beyond survey administration, generalizes better than earlier automated baselines, and exposes heterogeneity within the linguistic expression of identity fusion itself (Wright et al., 20 Sep 2025, Wright et al., 9 Feb 2026).