Example PhD

Information retrieval in folksonomies

Supervisor: Dr S. Schockaert

Keywords: Information retrieval, language modelling, logic engineering, folksonomies

Websites such as Flickr, Last.fm or YouTube allow users to organise content by attaching freely chosen, textual tags to resources of interest, resulting in an uncontrolled classification system called a folksonomy. These tags can then be used by search engines to find out which resources are most likely to be relevant to a given query. For instance, a search engine for Flickr photos can respond to the query "Millennium stadium" by simply listing all photos that have "millenniumstadium" or "millennium" and "stadium" as tags. However, as there is no obvious way to rank these photos, the quality of such a system may be quite variable: how do we know that the, say, top 10 photos shown to the user are the most interesting ones?

Since there is little evidence to judge which among a set of photos with tag "millenniumstadium" will best satisfy the user's request, one strategy might be to diversify the results as much as possible. We might, for example, diversify what is shown on the photo (e.g. the inside vs. the outside of the stadium), the context in which the photo was taken (e.g. at night vs. during the day, in winter vs. in summer), the kind of user that has taken the photo (e.g. a local vs. a tourist vs. a professional photographer), the camera settings (e.g. wide-angle vs. tele zoom), etc. For an ambiguous query such as "paris", we may moreover include photos that are relevant to its different senses (Paris, France vs. Paris, Texas vs. Paris Hilton). This problem of diversifying search results has recently gained considerable attention from the information retrieval community, although much of it has been focused on standard document retrieval. In folksonomies, however, different types of meta-data may be available (e.g. EXIF data for photos, user profiles, etc.) that should influence the diversification procedure in particular ways, i.e. algorithms that work well for retrieving diverse sets of Flickr photos may not help us to retrieve artists on Last.fm, although many of the underlying ingredients will be similar. This calls for diversification mechanisms that are sufficiently flexible and are easy to tweak, taking into account the nature of the available meta-data and the preferences of a given user. In this topic, the potential of fuzzy logics for this task will be assessed. Fuzzy logics are logics whose semantics are based on an infinite number of truth values, which enables them to model continuous domains. In particular, for the application of diversifying search results, fuzzy logics combine the flexibility of declarative, logic-based specifications, with the possibility of handling numerical features.

Diversification handles situations where there are many photos whose tag set encompasses all tags from the user's query. In contrast, there may also be situations in which there are too few photos that contain the exact query terms. We will focus in particular on queries that ask for photos of a given event, or of a given place. Examples of such queries include "Show me photos of waterfalls in Wales" and "Show me photos of the 2008 Olympics". To handle such requests in an adequate way, the system needs some understanding of the semantics of the events and place types that are being referred to. Such an understanding can be obtained by statistically analysing the tags that are attached to photos, in relation with their associated geographical location and time stamps (both of which are available for many Flickr photos). This topic will therefore also focus on statistical techniques to obtain semantic descriptions of photos, and the use of these descriptions for improving the performance of image retrieval systems.

Key Skills/Background: Open to computing graduates and postgraduates.

Contact: Dr S. Schockaert to discuss this research topic.