Research Programme

Field 1: Knowledge Processing

Focus 1: Knowledge Representation and Inference

Focus 2: Problem Solving and Action Planning

Focus 3: Acquisition and Communication of Knowledge

Introduction

Field 1: Knowledge Processing

Focus 1: Knowledge Representation and Inference

Ontologies for knowledge representation;  the ability to reuse, adapt and combine knowledge bases

Activities in this area serve the utilisation of varied ontologies (thesauri, classification systems, simple semantic networks like UMLS or WordNet, Penman Upper Model) and their mapping onto a common ontolingua. This will not only require the definition of representations between different languages but the closing of active knowledge gaps through automatic text-based learning processes as well. The main area of application is currently the field of medicine. The dissertations aim to improve  model-based re-usability and integration of knowledge descriptions in different languages and of different grains.

Spatial representation in the context of autonomous, mobile agents

One important ability of mobile agents is to determine  their own position, to recognise obstructions and other objects and to head towards target points. This necessarily requires the construction of an environment model and its use for navigating tasks. During model construction, an attempt must be made to create a correct model, despite any given uncertainties of the sensors. Here, deviations of the environment from the model should not lead to errors in orientation. A robust solution to this task can only be achieved by a combination of probabilistic and qualitative representation approaches.

Qualitative spatial and temporal representation and knowledge processing

Qualitative descriptions can be utilised if precise quantitative information on spatial or temporal configuration is not available or required. However, whether the qualitative formal systems on spatial and temporal inference suggested in the literature are indeed conceptually and inferentially cognitively adequate, which formal semantics they can be based on, which computability properties they have and what practical inference algorithms are (and to what extent these are valid as models of human inference in relation to spatial relations), is as yet only known to some extent and is currently being examined in projects of the SPP "Raumkognition" (spatial cognition).

Cased-based reasoning and representation of cases

CBR (Case-Based Reasoning) represents an alternative paradigm to rule-based forms of problem solving and planning. The examination of psychologically motivated forms of case representation, and in particular so-called "episodic representation", lies at the centre of these research interests. The utilisation of episodic knowledge in the analysis and diagnosis of problems and solutions and their practical application in knowledge-based instruction and learning systems, e.g. ELM, is of particular importance.

Intuitive and coherent evaluation of events

Content and context effects in  inferential thinking generally(Hans Spada) and  in spatial reasoning in particular(Gerhard Strube)

Solution types in deductive thinking tasks are heavily influenced by the contextual framework. These types of content and context effects are frequently used as evidence that humans do not think logically. The method of analysis  from the psychology of knowledge opens a new door to these questions. Initial empirical evidence and modelings based on Wason's selection task show that in connection with the content of a deductive task humans fall back heavily on their background knowledge and for the most part use it in a perfectly logical and correct way in inference. In this way, content knowledge and logical thinking should not be seen as opposites. The research project is based on recently completed dissertation (Beller). The "MeMoSpace" project deals especially with inference using spatial relations where content and context effects are also apparent and which are to be the focus of future research.

Children's thinking on their own knowledge and the knowledge of others

Modeling children's erroneous convictions gives a new perspective on current deficit hypotheses on children's thinking. Operational and representational deficits serve to explain evidence supporting the difficulty young children have in assigning the convictions they or other people had at an earlier time but which are in contrast to their current ones. This competence has received much attention due to its importance for the prediction and explanation of behaviour. Based on a recently completed dissertation (Wichmann), the aim will be to continue with the cognitive modeling of the operational and representational demands on children's thinking and to examine it empirically.
 

Focus 2: Problem Solving and Action Planning

"Intelligent" heuristic searching (incl. duplicate elimination), application in protocol verification

This research deals with the exploration of large state spaces, e.g. as they occur in NP-hard search problems in AI, for example in one person games, but also in the verification of communication protocols, in automated theorem proving and in robot control. Beginning with a start state, a goal state is sought. Transition rules prescribe how to get from one state to a possible successive state. This can be modeled with implicitly given graphs in which a search  for a start-goal-path is executed. Because it is normally possible to arrive at the same state along numerous possible paths, it is important to recognise so-called "duplicates" in good time so as to limit the search field. Therefore, we are particularly interested in duplicate elimination strategies and exploration methods which can function with a memory of a predetermined, definite size. Furthermore, we aim to examine strategies of  learning heuristics and strategies of active search field changes, e.g. through interaction with the environment.

Flexible availability of qualitative and quantitative knowledge in solving  physics tasks

Flexibility in problem solving and learning depends fundamentally on whether humans have different representations of the subject area at their disposal and whether they can apply these in a coordinated fashion. Based on a qualitative conceptual representation and a quantitative representation of school physics knowledge on classical mechanics, the aim will be to verify, using empirical  and cognitive modeling methods,  in which conditions extensive and flexible solutions are made possible, and in which positive learning results can be expected . The modeling of problem solving behaviour will also be used for the development of instructional and testing materials.

Goal management and action planning

Goal management is a central field of research within the context of "new" AI and cognitive science research focussed on cognitive systems. Action planning has become a great interest again in AI over the last few years due to the ability of recent planning algorithms to solve problems of a realistic size as well. Important research problems in these fields include the handling of multiple and frequently confligating goals, the further development of existing planning approaches (to enable the handling of planning formalisms using strong expressions as well), the handling of uncertain and dynamically changing information, in particular in dynamic environments with variable demands, for example in environments with several agents (multi-agent-systems).

The following research topics have been planned for this area:

• Heuristic searching in the context of uncertain information
• Heuristic searching and knowledge-based  problem solving in humans

Focus 3: Acquisition and Communication of Knowledge

• Support of individual learning processes using intelligent tutorial systems and
• Support of knowledge communication using "new media" (in particular multimedia systems)

Multimedia instruction and learning systems

This topic deals with the integration of live presentation and tele-presentation and the generation of a multimedia document for off-line use. One important tool in this respect would be a sufficiently powerful and comfortable whiteboard, to be used as an electronic substitute for blackboard and overhead projector, which would allow for the generation, loading, modification, selection and commenting of slides, for the loading of images, and for running and controlling animations and simulations. In order to enable tele-presentations the whiteboard will need to be multicast capable and the data streams it generates will need to be recorded and then processed in such a way to enable synchronisation of all data streams and optional access. In order to reach the appropriate functionality, the multimedia preparation of content and the development of corresponding tools should be pursued in parallel. In this context we are also interested how documents produced by the 'authoring on the fly' method can be integrated into instruction and learning systems and used in university lectures.
 

Knowledge acquisition and knowledge communication in cooperative and computer-based learning

Traditional instructional forms frequently result in an unsatisfactory intellectual comprehension and inadequate application of the subject matter. But how can learning environments be created so that learners can actively construct their own knowledge and apply it competently? The creation of cooperative computer-based environments is seen as one possibility. In this context the following questions arise, for example: how can cooperative learning forms and computer-based forms be connected and coordinated? Which role can people and computer-based components have in a cooperative, distributed learning environment? This project is associated with DFG projects (German Research Foundation) which are currently modeling problem solving processes in physics and information exchange in cooperative learning.

Automatic knowledge acquisition from texts; "intelligent" text mining and text classification in the filed of medicine

Activities in this field are aimed at the adaptation and extension of given domain knowledge bases using automatic learning from texts. Apart from methodological concerns (for instance a qualitative evaluation system for the credibility of learning hypotheses which has by now been expanded to include the administration of ambiguities), application aspects graduated according to demand strength are at the centre of our studies. This includes traditional information services like automatic text classification on the one hand, but increasingly also new applications like the "intelligent" text mining of facts, propositions and value judgements on the other.

The following research topics have been planned for this area:

• User-adaptive intelligent tutorial systems, tele-teaching with IST
• Computer-based training of mathematical proofs

Field 2: Language Processing

Syntactic analysis and verbal working memory in aphasics (Jürgen Dittmann) and in non-aphasics under memory load (Gerhard Strube)

Patients with pathological impairments in the verbal working memory resulting from brain damage always exhibit decreased performance in language production, sentence repetition, and sentence and text comprehension as well, whereby the relationship between these impairments is still unclear.

Projects - Dittmann:

• Which consequences does a relatively selective impairment in the verbal working memory have for the processing of subject/verb congruence in language production?
• Which consequences does such an impairment have for the reception of syntactically and semantically complex sentences, respectively?
• How is it possible to isolate the effect of an assumed working memory deficit on restrictions in processing syntactically complex sentences in aphasics?

Phonetic deviations in fluent and non-fluent aphasia

Within the framework of research on aphasic disorders in word production, a model representing word forms as scores of articulatory gestures has been formulated for German. Furthermore, in cooperation with the Dept. for Experimental Phonetics this model has been applied to the acustic-phonetic analysis of the single word utterances of non-aphasics. It has been shown that a complete and non-contradicting description of the articulatory variance in the utterances of non-aphasics as well as a complete error pattern in the aphasic utterances is possible using this "gestural" model. The application of the model to the conception of diagnostic tools and in a clinically orientated multimedia project on the use of PC-based learning environments in aphasia therapy is planned.

Analysis and modeling of human sentence processing

Within the framework of the SOUL projects, a theory of human syntax analysis and further semantic interpretation in the reception of linguistic utterances (including implemented computer models) was developed which is being extended in the current project, mainly with regard to the integration of thematic, morphological, syntactic and frequency information during reading. In particular, relative clause attachment preferences have proven to be a highly interesting area of study, which is being researched in cooperation with associates in England, Italy and Spain.

Machine text parsing of natural language

A performance grammar for German and its corresponding performance parser based on dependential constructs has been developed within the framework of the DFG programme "Cognitive Linguistics". Besides the integration of declarative and procedural knowledge in the context of a robust parsing process, this conception is also aimed at the immediate consideration of text structures in the grammar description and in the parsing process.

Non-standard problems of semantic text interpretation

Analyses in this field are aimed at dealing with illustrative language (mainly metonyms and their interaction with other, anaphoric text phenomena) and with the interpretation of value judgements using adjectives. It has been shown that the fundamental interpretation processes cannot be dealt with without extensive terminological modeling and classification-based reasoning.