Building Named Entity Recognition Models Efficiently Using NERDS
summary
An open source framework called NERDS wraps gazetteers, CRF, BiLSTM±CRF and ELMo models and exposes a fit/predict/save/load API; a fork fixes NLTK deps, adds minibatching and multi-entity support and simplifies the data format. Evaluation on the BioNLP 2004 set (≈500k train, 100k test) yields baselines: dictionary ≈0.65 F1, spaCy/neural ≈0.78 F1, ELMo ≈0.80 F1, ensemble ≈0.78–0.80. Known limits: serialization blocks parallel training; planned work adds scikit‑learn compatibility and BERT/Flair integration.
AI-written summary · about
chapters (17)
- 0:00Introduction
- 1:39Named entity recognition matters
- 3:25NER techniques overview
- 7:16Gazetteers and Aho-Corasick
- 8:20Conditional Random Fields
- 10:05Bi-directional LSTM for NER
- 11:51Word to character embeddings
- 13:38Contextual embeddings with ELMo
- 14:41Transformer models like BERT
- 17:09What is NERDS
- 18:12NERDS architecture and benefits
- 21:05Simplifying data formats
- 23:11Testing NERDS with bio NLP
- 24:54NERDS model improvements
- 27:27Comparing model performance
- 28:53Future work and API
- 30:40Q&A
related talks
49mNamed-Entity Recognition and Name Resolution Using Similarity Queries
22mEnterprise NLP with BERT | Optimized Text Analysis Framework
21mFrom 0 to a crowd sourced neural network in a jiffy
23mNeMo: Open Source Toolkit for Conversational AI
24mIs it possible to have entities within entities within entities?
40mPanel discussion "NLP"
34mExposing Dark Data in the enterprise with custom NLP
28mIndustry NLP: Multitask & Continual Learning