Conceptual Language Models for Context-Aware Text Retrieval

While participating in the HARD track our first question was, what an IR-application should look like that takes into account preference meta-data from the user, without the need of explicit (manual) meta-data tagging of the collection. Especially, we touch the question how contextual information can be described in an abstract model appropriate for the IR-task, which further allows improving and fine-tuning of the context representations by learning from the user. As a first result, we roughly sketch a system architecture and context representation based on statistical language models that fits well to the task of the HARD track. Furthermore, we discuss issues of ranking and score normalizations on this background.