# CORE-KG: An LLM-Driven Knowledge Graph Construction Framework for Human Smuggling Networks

**Authors:** Dipak Meher, Carlotta Domeniconi, Guadalupe Correa-Cabrera  
**arXiv:** [2506.21607](https://arxiv.org/abs/2506.21607)  
**Date:** 2025-06-20  
**Summary page:** https://www.summarizepaper.com/en/paper/2506.21607v1/

## Key points

- Human smuggling networks are complex and constantly evolving
- Legal case documents contain valuable insights but are often unstructured, dense with legal jargon, and filled with ambiguous references
- Existing methods for building knowledge graphs lack coreference resolution and often result in noisy or fragmented graphs
- A modular framework called CORE-KG has been proposed to address these issues
- CORE-KG utilizes a two-step pipeline: type-aware coreference resolution through structured prompts and entity/relationship extraction guided by domain-specific instructions
- Targeted preprocessing is applied to extract only the "Opinion" section from each case file to reduce noise and redundancy in constructing knowledge graphs
- In preliminary experiments using a sample of 20 cases, the system demonstrated its capacity to extract key entities and identify meaningful relationships
- The implementation of LLaMA 3.3 70B model for coreference resolution and KG construction ensures reproducibility and minimizes stochastic variation during inference
- The system operates within an open-source environment without relying on commercial APIs, enhancing accessibility and transparency
- Baseline evaluation against GraphRAG framework highlights the effectiveness of CORE-KG in reducing node duplication by 33.28% and legal noise by 38.37%
- By focusing on extracting relevant information from the Opinion section of legal documents through coreference resolution and structured entity/relationship extraction processes, CORE-KG offers a comprehensive approach to constructing interpretable KGs from complex narratives found in criminal cases involving human smuggling activities

## Summary

Human smuggling networks are complex and constantly evolving, making them challenging to analyze. Legal case documents contain valuable insights but are often unstructured, dense with legal jargon, and filled with ambiguous references, posing obstacles for automated knowledge graph (KG) construction. Existing methods for building KGs lack coreference resolution and often result in noisy or fragmented graphs. To address these issues, a modular framework called CORE-KG has been proposed. This framework utilizes a two-step pipeline: type-aware coreference resolution through structured prompts and entity/relationship extraction guided by domain-specific instructions. To reduce noise and redundancy in constructing KGs that capture the structure of criminal networks from legal texts, targeted preprocessing is applied to extract only the "Opinion" section from each case file. This section contains the main factual narrative detailing individuals involved, routes used, items transported, and events sequence relevant to building meaningful KGs. In preliminary experiments using a sample of 20 cases retrieved from the Nexis Uni database, the system demonstrated its capacity to extract key entities and identify meaningful relationships. The implementation of LLaMA 3.3 70B model for coreference resolution and KG construction ensures reproducibility and minimizes stochastic variation during inference. The system operates within an open-source environment without relying on commercial APIs, enhancing accessibility and transparency. Baseline evaluation against GraphRAG framework highlights the effectiveness of CORE-KG in reducing node duplication by 33.28% and legal noise by 38.37%, resulting in cleaner and more coherent graph structures suitable for analyzing criminal networks like human smuggling. By focusing on extracting relevant information from the Opinion section of legal documents through coreference resolution and structured entity/relationship extraction processes, CORE-KG offers a comprehensive approach to constructing interpretable KGs from complex narratives found in criminal cases involving human smuggling activities.