Unstructured legal documents such as court proceedings contain critical evidence about illicit networks, but extracting structured information from jargon-heavy text remains a challenge. While large language models (LLMs) can automate knowledge graph construction through named entity recognition and relation extraction (NER-RE), general-purpose models are not tailored to domain-specific entity and relationship definitions. A new pipeline called FineREX addresses this gap by fine-tuning a smaller LLM on a manually annotated dataset of 512 text chunks from human smuggling court cases.
| Metric | FineREX | General-purpose Baseline | Improvement |
|---|---|---|---|
| Entity F1-score | higher | lower | +15.50% absolute |
| Relation F1-score | higher | lower | +31.46% absolute |
| Legal noise (false positives) | nearly halved | standard | ~50% reduction |
| Node duplication rate (long docs) | 11.17% | 17.78% | -6.61 ppt |
| End-to-end processing time | reduced | standard | -50.0% |
FineREX Pipeline and Performance Metrics
According to the arXiv paper by Feldman, Elijah, Meher, Dipak, Domeniconi, and Carlotta (2026), FineREX is a streamlined knowledge graph construction pipeline built around a fine-tuned LLM for NER-RE. The model was trained on a manually annotated dataset of 512 text chunks extracted from human smuggling court proceedings. The authors report that FineREX achieves absolute improvements of 15.50% in entity F1-score and 31.46% in relation F1-score compared to a larger general-purpose LLM baseline. These gains translate into higher-quality knowledge graphs: legal noise from false positive extractions is reduced by nearly half, and node duplication on long documents drops from 17.78% to 11.17%.
Methodology: Fine-Tuning for Domain-Specific NER and RE
The core innovation of FineREX is its reliance on domain-specific fine-tuning rather than a larger general-purpose model. By focusing on a smaller, fine-tuned LLM, the pipeline avoids the need for document rewriting and redundant extraction stages. This design eliminates unnecessary computational overhead: the authors report an end-to-end processing time reduction of 50.0%. The manually annotated dataset ensures that entity types and relationship definitions align with the legal domain, capturing roles such as smuggler, facilitator, and victim, as well as relations like "organizes," "transports," and "recruits." This targeted approach reduces the noise that generic models introduce when parsing idiosyncratic legal language.
Domain-specific fine-tuning can substantially outperform larger general-purpose models while improving both the quality and efficiency of knowledge graph construction for illicit network analysis.
Implications for Illicit Network Analysis
The practical impact of FineREX extends beyond raw metric improvements. Lower node duplication means that knowledge graphs derived from long documents (e.g., multi-day court transcripts) contain fewer redundant entities, making it easier for analysts to trace connections between individuals and events. The halving of legal noise reduces the manual review burden on investigators, who currently must sift through false positive extractions. Furthermore, the 50% reduction in processing time allows analysts to iterate faster or scale analysis to larger document corpora without proportional compute cost increases.
While the current evaluation focuses on human smuggling, the methodology — domain-specific fine-tuning of a smaller LLM on a curated dataset with custom entity and relation definitions — is transferable to other jurisdictions or illicit activity types. The authors note that eliminating document rewriting and redundant extraction stages simplifies the pipeline and reduces latency, making it suitable for time-sensitive intelligence workflows.
For technology leaders overseeing compliance, due diligence, or trade security systems, the FineREX case demonstrates that targeted fine-tuning can deliver higher accuracy and lower operational cost than relying on general-purpose LLMs. As organizations increasingly adopt AI for document intelligence, the principle of domain-specific model adaptation may offer a repeatable pattern for extracting structured data from specialized, unstructured text.