Knowledge Enhanced Embodied Cognitive Interaction Technology

2008-03 till 2012-10
Research Areas: 

In the KnowCIT project we extend the conversational abilities of the conversational agent MAX by equipping him with access to collaboratively constructed knowledge drawn from the online encyclopedia Wikipedia. By means of the crowd-sourced knowledge resource, the agent is able to identify, label, track, and continue the topic of a dialog as the interlocutor of a human dialog partner. This allows him to answer questions, to detect topic changes and to react meaningfully to the challenge of dialogical dynamics.

Methods and Research Questions: 

In the project we aim to connect two yet unrelated areas of modeling cognitive systems with semantic-web technologies. This relates to distributed cognition as exemplified by rising web technologies – e.g. social ontologies – on the one hand and artificial cognition based on embodied virtual agents – e.g. MAX – on the other hand. Basic requirement for enabling a successful human-agent dialog is the access to conversational knowledge for both sides, the human and the artificial dialog partner. Unlike humans who are able to fall back on basic knowledge or natural language reference books (like Wikipedia), the computer-based opponent depends on an additional knowledge base representing this kind of basis in a machine-readable way. While information concerning language-processing needs to be assigned once only, e.g., by defining grammars or rules, knowledge about topics and their correlations has to be maintained and, if necessary, updated regularly. As many ambiguities might emerge in natural language conversations, solving these polysemies constitutes another important issue in human-agent interaction. The KnowCIT project aims to build interactive technology that enables artificial agents to explore crowd-sourced knowledge resources generated by large communities of web users. From a theoretical point of view we aim to tackle the grounding problem studied in cognitive science by interfacing artificial cognitive agents with social ontologies. That way artificial agents become beneficiaries of crowdsourcing so that their human users gain in turn from the increase of their communicative competence. This research is in the line of efforts to utilize social tagging systems such as, e.g., the Wikipedia, wikimanuals and other special wikis, which provide large resources of encyclopedic knowledge. In this context, we plan to exploit object knowledge as well as linguistic and metalinguistic knowledge (by example of so called wiktionaries) in a way that enables virtual agents to identify, label, track and to continue the topic of a dialogue in which they participate as the interlocutor of a human user. As a result virtual agents will be enabled to answer questions as well as to detect topic changes and to react meaningfully in this kind of dialogical dynamics. Consequently, to build more adequate models of thematic dialogue management in conjunction with their implementation and exhaustive empirical testing will be a major goal of the planned project. In summary, we aim at devising an enabling technology of all those cognitive interaction technologies, which integrate semantic means of interaction with human users.


A milestone of the KnowCIT project is the implementation of an open topic model that enables the artificial agent to identify and label the topic of a dialog to which he participates as an interlocutor of a human user. By interfacing the conceptual structures of Wikipedia, the project endows MAX with the capability to utilize more than 55,000 different thematic categories for the task of topic labeling. Another milestone of the KnowCIT project is its question-answering component. That is, KnowCIT enables MAX to exploit the document collection of Wikipedia by utilizing propositional information about more than 1.1 million entities for the task of knowledge representation and reasoning. Thereby, the topic model provides the basis to determine the set of topic-related patterns needed for hypothesis generation and answer candidate scoring.  The resultant question-answering component invites human dialog partners to ask natural language questions and to explore the encyclopedic knowledge of Wikipedia just by means of interacting with a charming conversational agent.