FineREX: Fine-Tuned NER-RE for Human Smuggling Knowledge Graphs

Research Approach

Constructing knowledge graphs from legal documents is essential for mapping human smuggling networks, but existing methods often rely on general-purpose large language models (LLMs) that struggle with domain-specific jargon and computational inefficiency. The authors introduce FineREX, a streamlined pipeline that replaces multi-stage, redundant extraction processes with a single-pass, fine-tuned model. By fine-tuning LLaMA 3.1 8B using a custom-annotated dataset of 512 text chunks, the authors optimized the model specifically for identifying entities (e.g., persons, locations, transportation means) and their relationships within the context of human smuggling.

Methodology

FineREX operates through four sequential phases: text chunking, fine-tuned NER-RE extraction, coreference mapping, and graph consolidation. Unlike previous frameworks that require expensive document rewriting and multiple extraction passes, FineREX uses a consolidation mechanism that applies coreference resolution directly to the initial extraction output. This approach resolves entity aliases (e.g., mapping various names to a single canonical entity) and merges duplicate nodes and relationships across different document segments. The model was trained using QLoRA (Quantized Low-Rank Adaptation) to achieve high performance with a smaller, more efficient parameter footprint.

FineREX demonstrated substantial improvements over a larger LLaMA 3.3 70B baseline, achieving absolute F1-score gains of 15.50% for entity extraction and 31.46% for relationship extraction. When tested on sixteen U.S. Department of Justice human smuggling cases, the framework reduced legal noise by nearly 50% and decreased node duplication from 17.78% to 11.17%. Furthermore, by eliminating redundant processing steps, the pipeline achieved a 50% reduction in end-to-end processing time, making it a more scalable solution for large-scale investigative analysis.

This research demonstrates that domain-specific fine-tuning is highly effective for specialized knowledge graph construction, even when using smaller, more efficient models. By improving the accuracy and efficiency of extracting intelligence from unstructured legal text, FineREX provides law enforcement and analysts with a more reliable tool for uncovering hidden network structures and disrupting illicit smuggling operations.