From Frozen Encoders to Multi-Agent LLM Pipelines: An Iterative Journey through Person-Place Relation Extraction in Multilingual Historical Texts
This paper documents the design decisions, failure modes, and key insights behind each strategy, including an honest analysis of the 17-point dev-to-test gap, providing a comprehensive account of iterative NLP system development under data scarcity, class imbalance, and multilingual historical OCR noise.