Named-Entity Recognition and Name Resolution Using Similarity Queries
summary
A pipeline processes 20,000 news items per day to extract named entities and resolve them against a database of 8 million companies to detect about 20 event types. Initial NER combines spaCy, CoreNLP and Stanza with a custom CRF model (~90% entity accuracy), with transformer-based models adding ~1.5% more. Name resolution replaces a large prefix tree with token-level vector indices (gensim) and uses cosine, soft‑cosine and Word Mover’s Distance for fast similarity queries.
AI-written summary · about
chapters (7)
- 0:00NER project intro
- 10:48Training own NER model
- 17:42NER machine learning
- 23:05Trie alternatives
- 30:36Name resolution with Similarity Queries
- 37:19Similarity Queries on keywords
- 42:01Q&A
related talks
35mBuilding Named Entity Recognition Models Efficiently Using NERDS
39mHerding Entities- Information Search and Synthesis in the Context of Transaction Data
27mWhat’s in a Name? – Fuzzy Matching Techniques for Proper Nouns
30mSemantic Search and Neural Information Retrieval
32mBuilding semantic search using Deep Learning
38mMaking neural content meaningful and truthful relying on OpenNLP tools
32mWhat is X up to? - NER and Relationship Extraction for Information Extraction
24mIs it possible to have entities within entities within entities?