Natural language processing (NLP) computationally analyzes unstructured human language to identify domain-relevant entities and relationships, supporting automatic knowledge graph construction. section 7.0
A conventional domain-specific NLP workflow commonly chains specialized models:
- Named entity recognition identifies mentions and classifies people, organizations, locations, or diseases. section 7.0
- Relation extraction, coreference resolution, entity resolution, and disambiguation add links and normalize references. section 5.3.4
- LLMs can replace the chained workflow through prompt engineering or fine-tuning, while traditional NLP and LLMs may coexist. section 5.4
Traditional NLP favors fast, inexpensive, CPU-capable, and isolated deployments but requires costly annotation, specialized expertise, and complex maintenance. LLMs reduce initial customization effort but must be weighed against concrete domain and business considerations. section 5.3.4 section 5.4
