A Set-Based Training Query Classification Approach for Twitter Search

Qingli Ma, Ben He, Jungang Xu, Bin Wang

Abstract: Learning to rank is a popular technique of building a ranking model for Twitter search by utilizing a rich list of features. As most learning to rank algorithms are supervised, their effectiveness is heavily affected by the quality of labeled training data. Selecting training queries with high quality is an important means to improving the effectiveness of ranking model for Twitter search. Existing approach for this problem learns a query quality classifier, which estimates the training query quality on a per query basis, but ignores the dependence between queries. This paper proposes a set-based training query classification approach that estimates a training query’s quality by taking its usefulness in combination with other training queries into consideration. Evaluation on standard TREC Microblog track test collection shows effective retrieval performance brought by the proposed approach.