Evaluating vector-space and probabilistic models for query to ad matching

In this work, we evaluate variants of several information re-trieval models from the classic BM25 model to Language Modeling approaches for retrieving relevant textual adver-tisements for Sponsored Search. Within the language mod-eling framework, we explore implicit query expansion via translation tables derived from multiple sources and pro-pose a novel method for directly estimating the probabil-ity that an advertisement is clicked for a given query. We also investigate explicit query expansion using regular web search results for sponsored search using the vector space framework. We find that web-based expansions result in significant improvement in Mean Average Precision