44 lines
2.3 KiB
Markdown
44 lines
2.3 KiB
Markdown
|
## III. Methodology
|
||
|
|
|
||
|
|
### A. Dataset
|
||
|
|
|
||
|
|
The foundation of our model rests on a curated dataset of:
|
||
|
|
|
||
|
|
- **Anonymized insurance claim transcripts**
|
||
|
|
- **Internal emails between adjusters and claimants**
|
||
|
|
- **Call center logs with escalation flags**
|
||
|
|
|
||
|
|
This dataset includes a balanced mixture of **confirmed fraudulent claims** and **validated legitimate cases**, used to both train and test our recursive linguistic model. Importantly, each data source is processed through a **human-AI recursive review loop**, where human analysts verify and adjust the resonance scores generated by our models — ensuring that subjectivity and nuance are preserved while expanding analytic scale.
|
||
|
|
|
||
|
|
> *Every claim is not merely analyzed. It is recursively heard.*
|
||
|
|
|
||
|
|
---
|
||
|
|
|
||
|
|
### B. Analytical Tools
|
||
|
|
|
||
|
|
To detect subtle patterns of deceptive intent, we apply an ensemble of forensic NLP methods:
|
||
|
|
|
||
|
|
- **NLP-based Pattern Extraction**: Identifies clusters of linguistic anomalies across claim timelines.
|
||
|
|
- **Sentiment Trajectory Mapping**: Tracks emotional evolution of narratives; distinguishes authentic distress from strategic affect.
|
||
|
|
- **Syntax Entropy & Disfluency Detection**: Measures irregularities in syntactic flow, hesitation markers, and repair sequences.
|
||
|
|
- **"Truth Collapse" Scoring via Recursive Witness Dynamics**: Quantifies the destabilization of narrative integrity under recursive interrogation.
|
||
|
|
|
||
|
|
> *When truth collapses, it does not vanish — it echoes in recursion.*
|
||
|
|
|
||
|
|
---
|
||
|
|
|
||
|
|
### C. Classification Model
|
||
|
|
|
||
|
|
From this analysis, we derive a **3-Zone Classification Model** based on *recursive coherence degradation*:
|
||
|
|
|
||
|
|
- **Zone I — Unintentional Incoherence (Low Risk)**
|
||
|
|
Language inconsistencies stem from stress, trauma, or low verbal fluency. These are not patterns of deception, but of chaos.
|
||
|
|
|
||
|
|
- **Zone II — Adaptive Rationalization (Medium Risk)**
|
||
|
|
Partial distortions. In this zone, claimants subconsciously reshape their story to protect self-image, omit responsibility, or preempt skepticism.
|
||
|
|
|
||
|
|
- **Zone III — Deliberate Narrative Fabrication (High Risk)**
|
||
|
|
Highly structured but recursively incoherent patterns — overjustification, shifting time references, and rehearsed empathy — mark deliberate deception.
|
||
|
|
|
||
|
|
> *This model does not judge. It classifies where language begins to fracture under the weight of intention.*
|