Daniel Shank, Data Scientist, Talla

on YouTube ↗17m

" summary

Goal is to detect repeated help questions and auto-serve past answers by matching new queries to historical chat lines. Uses active learning starting from about 100 seed labels and selects examples by uncertainty sampling (predictions near 0.5) to reach a usable classifier quickly while warning that label distribution skews. Builds representations with time-windowed word embeddings / matrix factorization tailored to problem contexts, notes topic models like LDA alpha sensitivity and that answer extraction remains a separate unsolved piece.

AI-written summary · about

" chapters (7)

1. 0:00Start
2. 0:11Daniel Shank
3. 0:38Smart Knowledge Management
4. 2:09Finding Issues and Matching Them to Incoming Questions
5. 6:58Active Learning
6. 11:42Dimensionality Reductions and Representations
7. 15:16Final Thoughts

" related talks

• 28mByron Galbraith, Chief Data Scientist, Talla at MLconf Seattle 2017Byron Galbraith · MLconf Seattle 2017
• 25mXavier Amatriain, VP of Engineering, Quora @ MLconf SEAXavier Amatriain · MLconf Seattle 2015
• 31mWhen humans teach machinesShai Hertz · WeAreDevelopers AI Congress Vienna 2018
• 22mDavid Talby, SVP Engineering, AtigeoDavid Talby · MLconf ATL 2015
• 38mFrom text to knowledge via ML algorithms - the Quora answerXavier Amatriain · Data By the Bay 2016
• 31mKeynote Q&A – Barbara PlankBarbara Plank · PyData Berlin 2017
• 37mActive Teaching, Human LearningVincent D Warmerdam · PyData Eindhoven 2023
• 28mFrom Tables to Answers: building QA System for In-Document SearchesVladimir Ageev · DSC Europe 23