A New Adaptive Indexing for Real-Time Web Search

Abstract

Adaptive indexing is an alternative to the self-tuning methods. It is especially useful in the scenario of unpredictable workload, and there is no idle time to invest in index creation. The authors present their ongoing work on a new realistic adaptive indexing that transforms the previous data crawling offline approach to a data-driven online approach. The proposed approach consists of three tasks: topic prediction, resource selection, and results combination and ranking. They work simultaneously to retrieve highly relevant results to the user's query in real time. To make the index highly refreshed and up-to-date, they collected data from highly prominent resources (e.g., Facebook, Twitter, Wikipedia, etc.). The empirical results showed that the proposed model is better than the traditional models that work offline and spend hours or days for building the index in different periods. In addition, the experiments showed that the training results are highly relevant for adhoc and diversity tasks.