Graph-based Change Detection for Natural Language

In an era of rapidly evolving text data, especially in specialized domains, tracking and analyzing changes is critical. This project leverages Entity-Relationship (ER) graphs to represent text structure, enabling improved detection of document-level changes. By transforming raw text into structured graphs, we enhance clarity, capture context, and identify key relationships.

Our approach involves multiple entity extraction techniques, from classical CRF models to neural methods, evaluated through ROUGE metrics for precision, recall, and F1-score. Current findings reveal areas for improvement in capturing multi-entity relationships and ensuring consistency, completeness, and correctness.

Future efforts aim to refine graph accuracy through fine-tuning, hybrid extraction techniques, and advanced graph construction methods. This work addresses critical challenges in knowledge graph generation, paving the way for better text analysis tools in domains like fraud detection, legal analysis, and more.

NeurIPS 2024 Poster can be found here: Poster