Leveraging Temporal Query-Term Dependency for Time-Aware Information Access

Abstract

Incorporating the temporal property of queries into time-aware information access methods has been shown to have a significant positive effect on a large number of search tasks, such as over microblogs and news archive. Recent work on time-aware search mostly rely on time-based relevance models that are built upon the language model framework. However, in this model, query terms are often assumed to be generated independently from each other. In this paper, we observe through a time series analysis that, query terms are temporally dependent and are frequently occurring within similar time periods when they deal with the same topics. In contrast to existing work, we propose a method that naturally extends the effective temporal language model and exploits this dependency at the term granularity level. Moreover, we reframe the task as a rank aggregation problem that fully exploits the temporal features of query terms. Experiments using the large-scale TREC Temporal Summarization 2013 and 2014 standard datasets empirically show that our method leads to significant performance improvements, when compared to state-of-the-art temporal ranking models.