Query performance prediction for microblog search: a preliminary study

Microblogging has recently become an integral part of the daily life of millions of people around the world. With a continuous flood of posts, microblogging services (e.g., Twitter) have to effectively handle millions of user queries that aim to search and follow recent developments of news or events. While predicting the quality of retrieved documents against search queries was extensively studied in domains such as the Web and news, the different nature of data and search task in microblogs triggers the need for re-visiting the problem in that context. In this work, we re-examined several state-of-the-art query performance predictors in the domain of microblog ad-hoc search using the two most-commonly used tweets collections with three different retrieval models that are used in microblog search. Our experiments showed that a temporal predictor was generally the best to fit the prediction task in the context of microblog search, indicating the importance of the temporal aspect in this task. The results also highlighted the need to either re-design some of the existing predictors or propose new ones to function effectively with different retrieval models that are used in our tested domain. Finally, our experiments on combining multiple predictors resulted in achieving considerable improvements in prediction quality over individual predictors, which confirmed the results reported in the literature but in different domains.