Creating Knowledgebases from unstructured text
summary
Claims that you can skip large labeled corpora by encoding domain heuristics into labeling rules and letting snorkel turn those rules plus ~100 human labels into probabilistic training data. Pipeline extracts candidates from text (e.g. FDA drug labels, clinical trials), iterates rule functions, and yields KB entries and relation labels; typical early results reach ~65–70% accuracy within weeks and paper-reports of ~0.8 F1 from ~110 seeds. Workflow uses reproducible tooling like a Docker container and succeeds when rules cover the key patterns.
AI-written summary · about
chapters (32)
- 0:00Accenture Technology Labs intro
- 1:47NLP project challenges
- 3:35Text extraction needs
- 4:39Text preprocessing limits
- 6:06Sentence extraction
- 7:30Training data scarcity
- 8:33Snorkel DeepDive minimal data
- 10:16Snorkel training sets
- 11:21Labeling rules with domain
- 12:26FDA drug label case study
- 13:50Clinical trial extraction
- 15:17Building knowledge bases
- 15:38Candidate extraction balance
- 17:23Snorkel feature extraction
- 18:04Rule functions design
- 19:30Evaluating labeling rules
- 20:12Training classifier minimal data
- 20:32Snorkel annotation UI
- 21:3765-70% accuracy minimal effort
- 22:19Docker for NLP workflows
- 23:03ML vs Snorkel pipelines
- 24:07Snorkel pros and cons
- 24:51Beyond medical text
- 25:34Snorkel code access
- 26:16Demo extraction annotation
- 28:24UI for annotation rules
- 30:08Hyperparameter tuning rules
- 32:22Extending Snorkel tasks
- 34:02Lemmatization vs stemming
- 35:46Labeling rules count
- 36:29Automated rule generation
- 37:17Q&A
related talks
30mEfficient data labelling with weak supervision
33mWeak supervision: a new paradigm for unreliable data
31mCreate text classifiers in a few hours using the open-source Label Sleuth system
51mHandling the problem of hand-labeled training data with data programming...
26mExtracting Medical Attributes from Clinical Trials
29mAnnotating data the right way
34mExposing Dark Data in the enterprise with custom NLP
33mBulk Labelling Techniques