A Supervised Learning Algorithm for Binary Domain Classification of Web Queries using SERPs
This addresses the issue of inefficient query routing for users of specialized digital libraries, though it is incremental as it builds on existing search engine infrastructure.
The paper tackled the problem of automatically detecting web queries that should be routed to local digital libraries instead of general search engines, by classifying queries as scholar or non-scholar domains using features from search engine result pages, achieving a precision of 0.809 and F-measure of 0.805.
General purpose Search Engines (SEs) crawl all domains (e.g., Sports, News, Entertainment) of the Web, but sometimes the informational need of a query is restricted to a particular domain (e.g., Medical). We leverage the work of SEs as part of our effort to route domain specific queries to local Digital Libraries (DLs). SEs are often used even if they are not the "best" source for certain types of queries. Rather than tell users to "use this DL for this kind of query", we intend to automatically detect when a query could be better served by a local DL (such as a private, access-controlled DL that is not crawlable via SEs). This is not an easy task because Web queries are short, ambiguous, and there is lack of quality labeled training data (or it is expensive to create). To detect queries that should be routed to local, specialized DLs, we first send the queries to Google and then examine the features in the resulting Search Engine Result Pages (SERPs), and then classify the query as belonging to either the scholar or non-scholar domain. Using 400,000 AOL queries for the non-scholar domain and 400,000 queries from the NASA Technical Report Server (NTRS) for the scholar domain, our classifier achieved a precision of 0.809 and F-measure of 0.805.