Dialogue Systems

A dialogue system —also called a conversational agent— is a program that can converse with people in natural language, whether in writing or by voice. Its aim is not merely to answer an isolated question but to sustain an exchange: to interpret what the user says, remember what came before, and produce a fitting reply at every turn.

These systems are one of the most visible applications of natural language processing (NLP) and a common form of interaction-focused AI agents. They are best understood through their purpose, their internal architecture, and their history.

Task-oriented or open-domain

The literature —such as the standard textbook Speech and Language Processing, by Daniel Jurafsky and James H. Martin— draws two broad families. Task-oriented systems pursue a specific, bounded goal: booking a flight, checking an order, resolving a customer-service request. They operate over a limited domain and are judged by whether the task gets done. Open-domain or chit-chat systems chase no closed goal: they talk about almost anything and are judged by the fluency and coherence of the exchange. Many real assistants blend both registers.

The classic pipeline: NLU, manager, NLG

For decades the dominant design was a pipeline of chained modules. First, natural language understanding (NLU) parses the user's turn to extract intent and the relevant entities (dates, cities, figures). Next, the dialogue manager tracks the state of the conversation —what is known and what is still missing— and applies a policy that decides the next move. Finally, natural language generation (NLG) turns that decision into a natural sentence. In voice systems this core is wrapped between speech recognition at the input and speech synthesis at the output. Rule-based approaches gave way to statistical ones, which learned those decisions from data.

From ELIZA to voice assistants and the LLM turn

The history includes ELIZA, the program Joseph Weizenbaum published from MIT in 1966. It analyzed sentences through decomposition rules triggered by keywords; the DOCTOR script made its replies resemble a psychotherapeutic interview. The paper itself notes that some users attributed understanding to the program even after its operation had been explained. That documents a tendency to project meaning, not that ELIZA understood or that every user believed a person was replying. Statistical systems and commercial voice assistants followed. The large-language-model turn widened the domain available through one interface: on November 30, 2022, OpenAI introduced ChatGPT as a research preview able to follow instructions and maintain conversational context. That did not unify the entire field. The Jurafsky and Martin textbook still distinguishes task-oriented from open-domain dialogue because completing a reservation under state and constraints is not evaluated like producing a fluent reply.

That leap does not solve everything. The NIST generative-AI profile treats “confabulation”—false or erroneous content presented confidently—as a risk to measure and manage. In long conversations, it also matters what information is retained, for how long, and with whose permission; fluency and memory are different properties. A paper on grounding in LLM-based dialogue systems explains why tying replies to events or external knowledge can reduce false or irrelevant information. For concrete tasks, the dialogue manager, state, confirmations, and tool validation remain explicit controls. None by itself guarantees a correct conversation from beginning to end.