Testing neural network, or how to convince AI love the humans
summary
Describes improving a neural-network–based NLP system used to build chatbots for US insurance companies, focusing on techniques the development team used to prevent inappropriate or trolling responses as the models became more capable. Mentions practical improvements to the product’s NLP pipeline and includes examples of problematic user–AI interactions that motivated those changes.
AI-written summary · about
related talks
20mHuman-Centered AI for Sustainable Development
48mThe Future Is Already Here, or How to Test Natural Language Processing Systems in Chatbots
18mDefending the unknown using AI and cybersecurity
15mPractical tactics to build reliable AI apps
38mAI-Powered Conversational Interfaces
03mDesigning for humanlike interaction
47mTesting Conversational AI
26mHow to develop useful and unbiased AI