Is it possible to have entities within entities within entities?
summary
spaCy’s open-source SpanCategorizer claims it can extract overlapping and arbitrary spans that NER cannot, replacing a brittle NER+text-classification pipeline for things like medical reviews. It pairs a suggestor (ngram, subtree, noun chunk or spin finder) with a classifier that pools span representations (first/last/mean/max) and outputs multi-label probabilities, reducing edge sensitivity. Implementations on review data and examples in the GitHub/ProjecTs repo show it works and improves precision/recall tradeoffs.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 1:08Named entity recognition problem
- 3:54Named entity recognition overview
- 6:43NER limitations in medical reviews
- 11:14Introducing SpanCategorizer
- 13:40Customizability in spaCy pipelines
- 15:46How SpanCategorizer works
- 19:18SpanCategorizer vs NER
- 20:23Resources
- 21:26Q&A
related talks
35mBuilding Named Entity Recognition Models Efficiently Using NERDS
49mNamed-Entity Recognition and Name Resolution Using Similarity Queries
37mBuilding a Scientific Taxonomy at Scale
31mCombining NLP with Structured Data
32mWhat is X up to? - NER and Relationship Extraction for Information Extraction
38mMaking neural content meaningful and truthful relying on OpenNLP tools
19mAddressing Treatment-Relevance in Biomedical Relation Extraction
39mHerding Entities- Information Search and Synthesis in the Context of Transaction Data