LLMs Underperform Graph-Based Parsers on Supervised Relation Extraction for Complex Graphs

This project is a scientific evaluation rather than a software product. Its primary value lies in the data and insights proving that specialized, smaller architectures (graph-based parsers) still outperform LLMs on high-complexity Relation Extraction (RE) tasks. With 4 forks in 8 days but 0 stars, it indicates early academic interest (likely peer researchers) rather than developer adoption. The defensibility is low because it is a reference implementation of a benchmark; the moat is the specialized knowledge, not the code itself. While frontier labs (OpenAI/Anthropic) are not building specialized graph parsers, they are rapidly improving the reasoning capabilities of LLMs, which may eventually close the performance gap identified here. This research serves as a 'reality check' for the industry trend of using LLMs for everything, highlighting specific niches where specialized NLP models remain superior. Its relevance is highest for teams building high-precision Knowledge Graphs for enterprise use cases where 95%+ accuracy on complex dependencies is required.