Cross-Document Temporal Relation Extraction with Temporal Anchoring Events


Automatically extracting a timeline on a certain topic from multiple documents has been a challenge in natural language processing, partly due to the difficulty of collecting large amounts of training data. In this work, we collect a dataset for cross-document timeline extraction from online news that gives access to metadata such as hyperlinks and publication dates. The metadata allows us to define a set of important events while linking them to time anchors, which opens the opportunity to scale up data collection. Furthermore, with this set of linked news articles, we propose a method to enhance the inference process of temporal relation prediction, by utilizing a model to link events to a set of anchoring events that are added to the inference program. We report performance of common neural models and show that our method can boost the performance of all baseline models.