A comparison of time-aware ranking methods

When searching a temporal document collection, e.g., news archives or blogs, the time dimension must be explicitly incorporated into a retrieval model in order to improve relevance ranking. Previous work has followed one of two main approaches: 1) a mixture model linearly combining textual similarity and temporal similarity, or 2) a probabilistic model generating a query from the textual and temporal part of a document independently. In this paper, we compare the effectiveness of different time-aware ranking methods by using a mixture model applied to all methods. Extensive evaluation is conducted using the New York Times Annotated Corpus, queries and relevance judgments obtained using the Amazon Mechanical Turk.