Exploiting Temporal Term Specificity Into a Probabilistic Ranking Model

Traditional information retrieval systems do not perform well in satisfying temporal users’ information needs. In this paper, we propose a temporal ranking model based on integrating temporal term specificity into a probabilistic ranking model for satisfying query time preferences. Implicit query time preferences are first derived using a temporal query profile, defined as being the distribution of the top k documents along a yearly-based time dimension. Temporal document ranking is based on calculating a temporal document score by combining term frequency and temporal term specificity. Temporal term specificity is derived from a temporal language model built per time partition, which refers to the word usage distribution in the documents annotated by a derived query time. The intuition is that the term usage distribution in documents tagged with specific time/date represents a clue to estimate the temporal relevancy of the document to this time. We conduct a user study on a portion of the New York Times corpus where results demonstrate that our approach outperforms the baseline search.