Computational Storytelling and Narrative Theory

In collaboration with the Institute for English and American Studies, this project attempts to conceptualize and implement a computational model of story generation, which is informed by narratological theory as well as cognitive modeling.

It approaches this problem starting from a mimetic stance towards fictional characters, that is, describing characters as intentional agents with discernible internal states like beliefs, desires, and affect. This allows investigating how narrative phenomena related to these paper beings can be computationally recreated in a multi-agent simulation system. Based on this internal perspective on narrative, the project also explores how from an external perspective the creative generation of plot can be controlled, and how the quality of the resulting plot can be evaluated, as a function of fictional characters.

The aim is to contribute to research on computational creativity by conceptualizing and implementing an evaluative storytelling system, and to narratology by proposing a generative narrative theory based on several post-structuralist descriptive ones. By using research methods from computer science to address problems from narrative theory this project also explores the use of interdisciplinary research methodology.

Further information
Leonid Berov

Concept Formation in Conceptual Spaces

In artificial intelligence, one can distinguish two layers of knowledge representation: Is the symbolic layer, abstract knowledge is represented in a structured, logic-based format, whereas in the subsymbolic layer, perceptual knowledge is stored in a numeric way, e.g., in the form of weights within a neural network. Ultimately, both approaches will have to be combined in order to arrive at a truly integrated system. It is however still unclear how exactly to accomplish this.

The cognitively inspired framework of conceptual spaces proposes to solve this problem by using an intermediate conceptual layer based on geometric representations: One can identify abstract symbols from the symbolic layer with regions in a high-dimensional space whose cognitively meaningful dimensions are based on subsymbolic perceptual processing.

In our lab, we explore this idea by formalizing the conceptual spaces framework in mathematical terms, implementing this formalization, and applying machine learning algorithms to it. By trying to extract meaningful dimensions from data sets, we aim to (partially) automate the construction of conceptual spaces. By grouping points in a conceptual space into regions, we aim to enable artificial systems to automatically discover new concepts based on unlabeled observations.

JESICS: Utilizing Cross-Domain Cognitive Mechanisms for Modeling Automatic Sketch Recognition

In collaboration with the Ain Shams University Cairo and German University Cairo, we investigate how humans learn the remarkable cognitive ability of “sketch recognition”. Humans can simply and efficiently learn from and recognize sketches without the use of complex reasoning. However, neither the recognition nor the production of sketches is easy to compute for machines. We claim that it is possible to computationally utilize cognitive mechanisms to enable artificial cognitive agents for demonstrating clear signs of human-like general intelligence, and we base our modeling on higher cognitive abilities by employing interdisciplinary research. The focus of the project employs cognitive science to fill an important gap in models of AI. This is a goal that needs first to consider an interdisciplinary way of exploring the main problem, formalizing its underlying theory, and modeling cognitively-inspired solutions to the problem.

Within the DAAD-founded project, the Osnabrück AI-group was involved in founding "The Joint Exploratory Society for Interdisciplinary and Computational Studies" (JESICS). This society constitutes the first Egyptian scientific community for studying, performing research in, and developing applications for interdisciplinary and cognitive sciences within artificial intelligent frameworks.

Conversational AI agents for collaborative e-learning: modeling conceptual knowledge and interaction around the concepts of multivocality and textual coherence to guide interventions

This research will be focused roughly between the areas of collaborative e-learning, the formation of conceptual knowledge and conversational AI agents. It will provide a new approach to modelling a process of conceptual knowledge building in collaborative interactions using indexes based on the concepts of multivocality and coherence. The models produced could serve to guide interventions of conversational agents or tutors during collaborative processes like problem solving in groups. Many of the methods used have already been described or implemented but will have to be adapted in innovative ways to model different stages during the collaborative knowledge building process.

Concept Invention Theory

The capacity of combinational creativity—i.e.when novel ideas are produced through unfamiliar combinations of familiar ideas—is difficult to recreate computationally. In particular, it is a hard task for autonomous computational systems to tackle the combinatorial explosion of potential combinations, and to be capable of recognizing the value of newly created ideas (concepts, theories, solutions, etc.), particularly when they are not specifically sought-this is the problem of creative serendipitous behaviour.

In COINVENT we aim to develop a computationally feasible, cognitively-inspired formal model of concept creation, drawing on Fauconnier and Turner's theory of conceptual blending, and grounding it on a sound mathematical theory of concepts.