In the KIRAS research project K.REX, a self-learning system is implemented for the content-related exploitation of evidence, which is intended to simulate the human ability to interpret documents by combining visual and textual aspects.
K.REX proposes an approach to capture elements of human perception in machine learned models to facilitate the interpretation of documents that otherwise cannot be considered in computer aided investigations and hence are lost from the chain of evidence. Methods for analysing text and images are combined in a multimodal system in order to increase the accuracy and pertinence of the computationally derived (possible) facts. In order to train the components of the system expert knowledge is efficiently recorded and formalised, specifically accounting for the idiosyncratic intricacies of each case and the ever changing patterns of fraud. Techniques for adaptive, dynamic learning will be explored to ensure fast and straightforward adaptation to new requirements with minimal effort for additional manual annotation.
The socio-technical implications of computer-aided investigations are substantial. Aspects of trust, accountability and transparency will be discussed and an ethical framework established. Early integration of sociological expertise can counteract and prevent effects like cognitive bias, preconceptions, or prejudice unintentionally instigated by the software. Iterative evaluation in a laboratory environment throughout the entire process ensures that the research meets the demands and expectations of the target audience.
Due to the expected reduction in turnaround times for the prosecution while improving the quality of the investigative results at the same time, K.REX has the potential to enhance the performance and efficiency of future analysis tools significantly.