Learn to Rank Tweets by Integrating Query-Specific Characteristics

Abstract

Current approaches to real-time Twitter search usually deploy a learning to rank (LTR) algorithm that incorporates diverse sources of tweet features, such as the content-based relevance score and the number of reposts, to find fresh and relevant tweets in response to user queries. In this paper, we argue that the user’s information need is a query-dependent notion, which is yet to be exploited for retrieval from tweets. To this end, we propose a learning to rank framework to improve the retrieval performance for real-time Twitter search by capturing the query-specific features. In particular, the proposed approach consists of two components: (1) a general ranking model learned from the training instances represented by features common to all queries; (2) a query-specific model learned by making use of LTR to select the most benefical expansion terms for each query. Finally, the query-specific model that reforms the original query topics with the extracted terms is combined with the general ranking model to produce the ranked list of documents in response to the given target query. Extensive experiments on the standard TREC Tweets11 collection show that the combined learning to rank approach outperforms the strong baseline, namely the conventional application of Ranking SVM.KeywordsReal-time Twitter searchQuery-specificLearning to rank