ADAM - Bootstrapping a Deep Neural Network Sequence Labeling Model with minimal labelling
summary
ADAM bootstraps NER for noisy product titles by generating weak labels from a sparse Knowledge Graph, matching ~1M stock items and retaining 8,000 weak labels after quality checks, training a domain word embedding on ~13M titles, then a Bi-LSTM + CRF sequence model and an active learning loop that samples 1,000 samples per iteration via extrinsic sampling. Baselines start at ~22% full-sequence match (brand 32%, category 36%); after four iterations full-sequence rises to ~82% and per-attribute accuracy climbs to roughly high 80s.
AI-written summary · about
chapters (14)
- 0:00Intro
- 3:26Why is it important
- 5:13Previous approaches
- 11:48Architecture
- 12:38Previous Approach
- 14:00ADAM Architecture
- 20:32Loss Function
- 23:36Results
- 24:48Baseline Results
- 25:59Extrinsic Sampling
- 28:22Gaming Clustering
- 30:38Positive Results
- 32:27Contribution
- 34:21Future Scope
related talks
51mEfficient Deep Learning with Humans in the Loop
33mBulk Labelling Techniques
30mStarting deep learning projects without enough labelled data
23mLearning with Limited Labels & Weak Supervision
20mPragmatic Deep Learning for image labelling
30mActive Learning and Human-in-the-Loop Machine Learning
45mActive learning and weak supervision in NLP projects
30mEfficient data labelling with weak supervision