Titel: Evaluating and Predicting Timelines

Abstract: Timeline summarization (TLS) is the task of automatically constructing timelines for events, employing large corpora.  Due to the temporal dimension and the scale of the task, TLS poses unique challenges that are not encountered in standard single- or multi-document summarization.

In this presentation, we present recent and ongoing work on two challenges.  In the first part of the talk we consider how to adequately evaluate timelines.  Previous work applied standard summarization evaluation metrics without considering the temporal dimension.  We devise a family of temporally sensitive evaluation metrics and evaluate these theoretically and empirically. 

In the second part of the talk we devise an approach to TLS that casts the problem as optimization of a submodular function under constraints.  Due to more complex constraints this goes beyond a similar effort in multi-document summarization. We analyze the effect of different constraints and objective functions.