How we trained a bot to understand fashion requests
summary
Chatshopper builds a conversational commerce layer for fashion shops and replaces a human WhatsApp service with an automated bot that went live in 2016. Natural language work uses 10,000 labeled expressions to classify 41 intents via two models: a linear SVM on summed 300‑dim word embeddings (≈93% accuracy) and a bi‑directional LSTM RNN (≈95% accuracy). Named entities use a conditional random field with about 90% entity recall for category/color/price, plus image neural nets to enrich product metadata for better search.
AI-written summary · about
chapters (12)
- 0:00Introduction
- 6:26Intent Classifications
- 7:10Machine Learning
- 8:46Word Embedding
- 13:53Support Vector Machine
- 16:48Recurrent Neural Network
- 19:08RNN vs SVM
- 20:28Named Entity Recognition
- 22:08Hidden Markov Model
- 26:08Features
- 28:01Performance
- 29:13Conclusion
related talks
35mDesigning Experiences for Chatbots
16mSolve for AI
47mTesting Conversational AI
35mNatural and conversational shopping experiences
37mLeveraging Generative AI for Enhanced E-commerce
27mA primer to write a conversational chatbot
24mConversational Interfaces for Chatbots and Artificial Intelligence
01h 04mBuilding Expertise Into Generated Conversations