Main concepts

Participants

Agents and users are the participants in a dialogue. Generally, it is assumed that the agent is a conversational system and the user is a human. However, the agent might be played by a human (“Wizard-of-Oz”) and the human user might be simulated.

Dialogue

A dialogue is turn-based and takes place between two participants (agent and user) that converse about a topic, depending on the purpose and scope of the conversational system (e.g., open-domain information seeking or movie recommendations). A dialogue is composed of a list of utterances (that may or may not be annotated) and may be accompanied by additional metadata.

Dialogue Orchestration

The dialogue orchestration includes the dialogue connector and the platform that enable the participants to communicate. The dialogue connector manages the dialogue between the participants, while the platform ensures that the participants can read and answer each other’s messages.

Natural Language Understanding (NLU)

NLU is responsible for obtaining a structured representation of raw text utterances. Currently, it entails dialogue acts recognition, key-value annotation (e.g., user sentiment), and user satisfaction prediction.

For example, in case of restaurant recommendation, the dialogue act DISCLOSE(food=Turkish,price=cheap) can be extracted from the utterance “I am looking for a cheap Turkish restaurant.”

Natural Language Generation (NLG)

NLG is responsible for creating raw text utterances from their structured representation.

For example, in the case of movie recommendation, “I recommend the movie Titanic.” can be generated for the intent RECOMMEND and slot value pair (title, Titanic).