Conversational Agents and Dialogue Systems
Dialogue systems, chatbot architectures, dialogue act tagging, slot filling, seq2seq models, RAG for dialogue, and task-oriented agents.
Summary
Building systems that hold meaningful conversations. From foundations and dialogue act tagging through seq2seq response generation, retrieval-augmented generation (RAG) for factual dialogue, and task-oriented agents (e.g. restaurant booking with Ollama + Llama).
Topics & Practice
Week 1: Foundations of Dialogue Systems
Introduction to dialogue systems: types (chitchat vs task-oriented, retrieval vs generative), dialogue acts, pipelines (NLU, dialogue state, NLG), and evaluation. Sets the stage for the labs that follow.
Week 2: Dialogue Act Tagging
Supervised dialogue act classification using the Switchboard Dialog Act Corpus. Train and compare DA tagging models as a foundation for intent-aware dialogue systems.
Week 3: Seq2Seq for Dialogue
End-to-end generative dialogue with sequence-to-sequence models. Build a seq2seq model for open-domain conversation and compare with retrieval-based baselines.
Week 4: RAG for Dialogue
Retrieval-augmented generation for information-seeking dialogue. Use Contriever to retrieve evidence and condition LLM responses on it to reduce hallucinations.
Week 5: Task-Oriented Agent (Restaurant Booking)
Build a prompt-based restaurant booking agent with Ollama and Llama 3.1. Use prompts and simple APIs to guide the model to call the right functions and complete bookings.