Estimating Event Focus Time Using Neural Word Embeddings

Time associated with news events has been leveraged as a complementary dimension to text in several applications such as temporal information retrieval, news event linking, etc. Short textual event descriptions (e.g., single sentences) are prevalent in web documents (also considered as inputs in the above applications) and often lack explicit temporal expressions for grounding them to a precise time period. For example, the event description, "France swears in Emmanuel Macron as the 25th President", lacks temporal cues to indicate that the event occurred in the year "2017". Thus, we address the problem of estimating event focus time defined as a time interval with maximum association thereby indicating its occurrence period. We propose several estimators that leverage distributional event and time representations learned from large external document collections by adapting the word2vec paradigm. Extensive experiments using two real-world datasets and 100 Wikipedia events show that our method outperforms several state-of-the-art baselines.