What is Task-Oriented Dialogue?

Examples:

Understand user utterances to perform tasks like provide info, manipulate objects, navigation
Generating response to user requests in a consistent and coherent manner while maintaining dialogue history
Managing user and system turns effectively and recovering from errors from potential misunderstanding of user utterances
Engaging user with adaptive and interesting conversations while progressing towards the common goal

The dialogue context is important, and we need to properly capture the back-and-forth to get all necessary info for the common goal.

Why is Task-Oriented Dialogue Challenging?

Context of dialogue history is paramount to achieve task success
Ability to capture back-and-forth interaction, estimate belief state of user goals and resolve anaphora during discourse
Piece together all relevant and necessary info to complete the task
Domain specific nature of dialogue makes it harder for adaptation and generalization
API calls to web-services may change over time and require frequent maintenance of dialogue managers

Task dialogue consists of:

Turn taking
Speech acts
Grounding

Task-Oriented Dialogue Characteristics

Turn Taking

Dialogue is often interleaved between at least 2 speakers (incl. agent)
Speakers holds the floor for a given time and takes opportunity to speak/inform/request

Speech Acts

Assertives: Speaker suggests or puts forward info to the agent
Directives: Speaker asks or requests info from agent
Commissives: Speaker agrees or opposes agent (yes/no)
Expressives: Greets the agent (hello/goodbye)

Grounding

Whether the user and agent mutually believe or agree on set of info exchanged
Agent or user acknowledge each other’s info

System

Based on a domain, the user might have an intent. To fulfill the intent, the information from the slots are needed.

Terminology

From Domains we can get Intents
Slots are spans contained in a given utterance that leads to a certain intent
- E.g., if we want to book a flight, the slots are the necessary info for booking that flight

<aside> 📌 SUMMARY: Task oriented dialogue involves having a user interact with the system to achieve something. This is challenging as we interact in natural language, and the system has to correctly identify the context to fulfill the user’s goal.

</aside>

Date: October 16, 2025

Topic: Automatic Speech Recognition

<aside> 📌 SUMMARY: ASR models translate utterances into words that a language model can understand, and is evaluated by Slot Error Rate

</aside>