Many document collections are by nature dynamic, evolving as the topics or events they describe change. The goal of temporal text mining is to discover bursty patterns and to identify and highlight these changes to better enable readers to track stories. Here, we focus on the news domain, where the changes revolve around novel, previously unpublished, "facts" that have an effect on the story developments. However, despite intense research activities on bursty patterns, a lack of common procedures today makes it impossible to compare methods in a principled way. To close this gap, we (a) investigate how different temporal text mining methods discover novel facts and (b) present an evaluation framework for methods assessment, consisting of a set of procedures and metrics for cross-evaluating models. Bursty patterns are transformed into queries for sentence retrieval, either with or without taking into account internal pattern structure, and these sentences are compared with a set of editor-selected ground-truth reference sentences. Our experiments on different classes of temporal text mining show that different methods perform at similar levels overall, but provide distinctive advantages in some settings. The experiments also demonstrate the benefits of using patterns' internal structure for query generation.
展开▼
