Mining multimedia subjective feedback

Personalising web search engines, a crucial issue nowadays, would obviously benefit from the system's ability to capture such an important aspect of a user's personality as visual impressions and their communication. Unfortunately, such impressions are intrinsically variable and very difficult to make explicit. In this paper, we propose a methodology to dynamically create personalised models of visual impression words. We suggest that it can be achieved by exploiting multimedia data mining on the large amount of information contained in an image and in its mapping onto an impression word. The continuous adaptation of each model utilises not only relevance feedback but also externalisation feedback. The externalisation process, triggered by the data mining activity, is supported by a conceptual space whose tools allow users to transform their mental model into formal specifications. For experimental validation purposes, a web-based meta search engine has been developed that can retrieve images on the basis of impression words.