Building a Scientific Taxonomy at Scale
summary
Pipeline builds a 115K-concept layer attached to OpenAlex and claims per-field graph clustering at scale using SPECTER2 embeddings, mutual kNN graphs, and Leiden with CPM resolution to cluster 100K+ nodes; Qdrant handles hierarchical vector attachment and Azure OpenAI (plus self-hosted models) runs five LLM stages for filtering, classification, renaming, explanations, and validation. Retriever returns ~150 candidates per paper across eight text-splitting views, deterministic pruning reduces to ~100, and an LLM reranker selects the final 5–8 concepts for tagging. Tools: SPECTER2 + Leiden (CPM) on 100K+ nodes.
AI-written summary · about
chapters (8)
- 0:00Scientific taxonomy project
- 2:53Why keywords are not sufficient for taxonomy
- 4:36Fifth taxonomy layer
- 15:56Concept hierarchy
- 20:15Tagging papers
- 32:36LLM choices for tagging
- 34:42Model performance
- 36:06Lessons and future
related talks
33mDeep research at scale
18mData Classification at Scale: Taming the Hydra
13mBuilding High-Performance AI Literature Processing
28mSearching for Research Fraud in OpenAlex with Graph Data Science
31mWhen humans teach machines
39mUnsupervised NLP Classification with Clustering
31mBuilding a Scalable, Reproducible Ontology System to Improve Deep Learning Data
30mProductionizing an unsupervised machine learning model to...