Inferring local synonyms for improving keyword suggestion in an on-line advertisement system

In this paper we present a keyword suggestion mechanism for supporting advertisers wishing to publish ads in content-targeted advertisement systems. The method infers "synonymy" between keywords by mining a database of previously submitted ads, and uses such information for suggesting relevant and non-obvious keywords to advertisers. Automatic word-sense disambiguation is provided implicitly by our keyword ranking procedure. We perform on-line comparison of our method with another keyword suggestion system being currently used in the largest Portuguese web advertisement broker with presence in four different countries, by redirecting 50% of keyword suggestion requests to each of the two systems. We propose a novel set of evaluation measures to compare the performance of keyword suggestion systems in such experimental setting. Results show that ads containing keywords suggested by the method we propose are selected for being printed more frequently at similar, or often superior, click-through rates. This results in a potential global revenue increase for the ad broker.