Social bookmark (tag) systems such as deliciou.com suffer a problem of increasing “tag noise” as more people join in and contribute tags. This was first noted in a paper by Ed Chi and Todd Mytkowitcz in which they applied information theory to measure this phenomenon., The plot on the left of the Figure above comes from some of their work. It shows that the mutual information between tags and documents decreased over time in delicious.com (a finding that has been replicated for several other systems). What this means is that it takes more tags to specify and search content.

The MrTaggy system (to be released soon) developed in the Augmented Social Cognition Area at PARC is one attempt to solve the problem. It does this through a combination of intelligent back-end algorithms and a user interface that promotes exploratory learning so that people can rapidly learn to communicate their information needs more precisely.

We recently conducted a study to demonstrate this effect.  One group of users worked with a “Base” version of MrTaggy that was just a search page that would return a list f search results like most standard search engines. The “Exploratory” version presents used with the search results plus associated and related tags, and the interface allowed the user to provide relevance feedback to the system by clicking on tags that seemed related to their information goal. The two groups worked with multiple queries in three different task domains and were tested in a variety of ways.

One of the tests required users to come up with keywords that could be used in the task domains they had been working on. The users in the full condition produced significantly more  (see left). We also statistically analyzed the influence of pre-experimental background knowledge (in the three domains they worked on). Background knowledge was correlated with the production of valid keywords for the Baseline users, but this correlation  disappeared for the Exploratory users. With the help of the Exploratory MrTaggy interface, low-knowledge users became more like high-knowledge users.

Summary

The MrTaggy tag search engine employs an exploratory search interface that provides users with a ways of seeing related tags and allows them to provide relevance feedback to the system to narrow down their results sets. Interacting with this system appears to help users learn the keyword vocabulary for subject domains.