This paper describes the prototype interactive search system being developed within the AutoAdapt project1. The AutoAdapt project seeks to enhance the user experience in searching for information and navigating within selected domain collections by providing structured representations of domain knowledge to be directly explored, logged, adapted and updated to re ect user needs. We propose that this structure is a valuable stepping-stone in context-rich logging of user activities within the information seeking environment. Here we describe the primary components that have been implemented and the user interactions that it will support.
Searches within document collections like intranets di er
from those within the general World Wide Web [
One of the di culties in using traditional logging of user activity, such as submitted query terms, URL clicks, and page viewing time, to adapt search systems is the lack of su cient context for identifying the user actions that are truly relevant to the user's information need. We implement methods of explicitly visualising domain models to accompany each search step, in addition to a list of links to search results, and a set of query term suggestions. By concurrently logging user interaction with the these components we have a mechanism to enable subsequent weblog analysis. For example, di erent document selections following the exploration of the same path may indicate relevance between documents, di erent paths leading to the same document may indicate relationships between paths, a comparison of path before and after a document selection should yield some understanding of the nature of the document selection. We present here a working system including a graphical domain model presentation, a document list and term suggestions designed to capture the described information.
It is frequently pointed out that users are reluctant to leave
any explicit feedback when they search a document
collection. However, implicit feedback, e.g. the analysis of log
records, has been shown to be good at approximating
explicit feedback. For example, users often reformulate their
query and such patterns can help in learning an improved
ranking function [
In gure 1 we can see a screenshot of our demonstration system. There are four basic components, a) a simple entry box for query terms, b) a list of URLs with associated snippets, c) a graph displaying a segment of a domain model, and d) a list of suggested terms for query re nement.
The user enters a set of query terms, this results in a number of documents being returned. Using the query terms, additional terms are automatically extracted, e.g., from the domain model and the highest ranked documents. These terms are then represented as nodes in a graph as a segment of the domain model. The user can then traverse the graph by clicking on the nodes (the e ect is to make that term the centre of the graph). On term selection the list of suggested terms is updated to show terms closely related to the selected term. The user can then add the term to the existing query or use it as a new query.
The modular nature of the software allows a standard user interface and logging structure (described in the next section) irrespective of the domain model creation and adaptation algorithms employed. We can, for example, examine di erent interaction styles and evaluate other domain model visualisation tools.
In addition to logging user query terms with presented and selected URLs it was decided to log the segment of the domain model presented to the user. As we intend to modify the domain model over time based on responses to the model presented, it is essential that a complete copy of the presented model segment is retained in the database. Of particular interest is the term positioned at the centre of the graph and the co-ordinates of the other terms. Using this information and the term clicks we can determine how the model was traversed, allowing us to identify which terms were also visible and ignored. Suggested terms (derived by the model) are also recorded along with any selection (to expand, or replace initially submitted query terms). The logging structure allows us to record a number of user decisions without the need for explicit feedback. For example, the selection of a term in a domain model can provide a ranking of terms, i.e., above those shown but not selected. Also, suggested terms derived from a particular traversal can be ordered. In addition, we can compare sessions that have resulted in the same URL being selected in order to capture related terms or similar portions of the domain model. It is also possible to compare portions of di erent domain models to discover missed relationships or terms.
As the next step, we propose to test the infrastructure in this document across several domain collections and model creation/adaptation algorithms to extensively evaluate the e ectiveness of the system in capturing the context of user interaction.
AutoAdapt is funded by EPSRC grants EP/F035357/1 and EP/F035705/1. The JIT visualisation toolkit2was used for the domain model visualisation. 2http://blog.thejit.org/javascript-informationvisualization-toolkit-jit Figure 1: Screenshot of AutoAdapt Demo System.