Automated Interactive Domain-Specific Conversational Agents that Understand Huma...
summary
AutoConcierge maps utterances to structured predicates and back to sentences: uses GPT-3 to convert dialogs into predicates then runs those predicates through s(CASP) goal-directed ASP for commonsense reasoning to check consistency, ask for missing slots, and compute recommendations. Demo uses a 150–160 restaurant dataset around UT Dallas, reports predicate extraction ~89% with GPT-3 and ~92% with GPT-4, outperforms end-to-end transformers on coherence and factuality, and limits include narrow vocabulary and limited scalability to broader domains.
AI-written summary · about
chapters (14)
- 0:00Domain-specific conversational agents
- 1:25Human-like chatbot design
- 3:17STAR framework
- 6:34Knowledge extraction with LLMs
- 9:26SCASP knowledge processing
- 11:11AutocCon architecture
- 14:10Vocabulary for reasoning
- 17:39CT state machine
- 20:57Live AutocCon demo
- 34:20Demo walkthrough
- 40:24Hallucination in responses
- 42:14Evaluation results
- 44:43Limitations and scalability
- 45:28Conclusion
related talks
31mGenerative Conversational AI Interoperability
01h 04mBuilding Expertise Into Generated Conversations
25mMultimodal Conversational Interfaces with GPT and Vision AI
58mA Primer on AI Powered Conversational Experiences and Language Models
34mBuilding Multi-Agent Generative-AI Applications with AutoGen
10mKeynote: Next-Gen AI Applications via Multi-Agent Conversation
40mAI Agents in Practice
45mWeb Development in the Age of AI