Motivated by the observation that expert searchers follow well-defined search procedures in order to obtain comprehensive information on that Web, in our prior work, we developed a prototype domain portal called the Strategy Hub to provide these search procedures to novice searchers. The search procedures in the prototype were entirely handcrafted by search experts, making further expansion of the Strategy Hub cost-prohibitive. However, a recent study on the distribution of healthcare information on the web suggested that search procedures can be automatically generated from pages that have been rated based on the extent to which they cover facts relevant to a topic.
This paper presents the results of experiments designed to automate the process of rating the extent to which a page covers relevant facts. To automatically generate these ratings, we used two natural language systems, Latent Semantic Analysis and MEAD, to compute the similarity between sentences on the page and each fact. We then used an algorithm to convert these similarity scores to a single rating that represents the extent to which the page covered each fact.
These automatic ratings are compared with manual ratings using inter-rater reliability statistics. Analysis of these statistics reveals the strengths and weaknesses of each tool, and suggests avenues for improvement.