WP6: Adaptive situated dialogue processing
Situated dialogue is a means for a robot to extend or refine knowledge about the environment. For this, the robot needs to be able to establish with a human some form of mutually agreed-upon understanding — they need to reach a common ground. The goal of this WP is to develop adaptive mechanisms for situated dialogue processing, to enable a robot to establish such common ground. We will focus on dialogues for continuous learning. In continuous learning, the robot is ultimately driven by its own curiosity, rather than by extrinsic motivations. Therefore, we want to conceive of dialogue as peer-to-peer, mixed-initiative communication — either the robot or the human can be the one to be asked to clarify, explain, or perform something.
Establishing common ground requires the robot to be able to process (i.e. comprehend and generate) clarification requests and explanations. A clarification request is a request for information to help overcome a breakdown in communication, or in understanding a situation. A request can be a single utterance, but also an entire sub-dialogue — e.g., when a request needs to be refined, or rephrased. An explanation provides information about why an agent does something, or believes a certain fact to be true. To achieve common ground in dialogue for continuous learning, these strategies are to serve two related purposes: They help the robot and the human to achieve transparency in what the robot needs to learn about the situation, and then interactively set up an appropriate context (scaffolding) in which a clarification request, an explanation, or a task can be used to trigger an appropriate learning goal.
This problem is challenging because the robot’s knowledge is continuously being adapted and extended — and this requires dialogue processing to be adaptive and extendible, too. First, as categorical knowledge grows, the robot needs to learn online how to use context information to focus clarification and explanation on what is relevant. Categorical knowledge is an associative network of concepts, and only some may be relevant for the robot to talk about to get the answer it seeks. Second, because the robot acquires knowledge over a period of time with varying degrees of supervision/autonomy, we cannot assume that the robot’s grammatical competence always has the adequate coverage to express newly acquired (categorical) knowledge. This means the robot needs an ability to learn how to linguistically convey new categorical knowledge.
The objectives of WP6 are as follows:
- Provide the robot with the basis for interactively establishing a mutually agreed upon common ground of the user and the robot.
- Provide verbalisation of the basic and combined perception-based concepts learned in WP5 (extension of lexico-grammatical knowledge).
- Use motivation systems to guide mixed-initiative communication.
- Investigate to what degree situated dialogue can make the robot transparent to the user in terms of intentions, knowledge level and other internal properties.
- Active scene manipulation to create situations that support the communication of certain aspects of knowledge (e.g. to illustrate what a certain concept refers to).