Byron Galbraith, Chief Data Scientist, Talla at MLconf Seattle 2017
summary
A production conversational service desk sits on three pieces: a conversational knowledge base, a lightweight ticketing system and intelligent workflows, built to answer employee Q&A and fall back to humans when needed. Experiments show pure embeddings hurt ad hoc IR but combining BM25 + embeddings improves retrieval; rare out‑of‑vocabulary tokens remain discriminative so an ensemble with TF‑IDF is used. System surfaces up to four suggested questions, supports real‑time one‑shot learning via fast linear models and a cascade retrain, and keeps a human‑in‑the‑loop edit UI so new answers appear almost immediately.
AI-written summary · about
chapters (9)
- 0:00Start
- 0:08Neural Information Retrieval
- 0:21Intelligent Conversational Service Desk
- 1:11Chatbots
- 16:31Operationalizing Deep Learning
- 22:06Expectation Management
- 23:00Summary
- 23:52Questions
- 24:03How Do We Represent Context in those Conversations
related talks
17mDaniel Shank, Data Scientist, Talla
31mMachine Learning-Enabled Chatbots
46mHandling Support Tickets with AI Without Burning the Planet
41mConversational Agents at Scale: Retrieval and Generative approaches
31mTaking AI/ML to the next level: Text!
26mAdvancing Multi-Modal Search Capabilities in Search Pipeline
50mCreating a scalable (...) with Elixir, Python & Tensorflow
25mXavier Amatriain, VP of Engineering, Quora @ MLconf SEA