Dialogue systems let users interact in natural language to get information, complete tasks, or have open-ended conversation. They fall into broad categories: chitchat (social, open-domain) versus task-oriented (booking, support, form-filling); retrieval-based (pick a predefined response) versus generative (produce new text, e.g. with seq2seq or LLMs).
Dialogue act (DA) tagging classifies each utterance by its function — question, statement, acknowledgment, request — and is a core step in many pipelines. A typical pipeline has NLU (understand intent and slots), dialogue state tracking (what the user has asked for so far), policy (what to do next), and NLG (surface a response).
In the following weeks you will implement dialogue act tagging on the Switchboard corpus, build a seq2seq generative model for open dialogue, add retrieval-augmented generation (RAG) to ground responses in external knowledge, and finally build a task-oriented restaurant-booking agent with local LLMs (Ollama + Llama) and simple function calling.