Redundancy reduction for multi-document summaries using A* search and discriminative training

In this paper we address the problem of optimizing global multidocument summary quality using A* search and discriminative training. Different search strategies have been investigated to find the globally best summary. In them the search is usually guided by an existing prediction model which can distinguish between good and bad summaries. However, this is problematic because the model is not trained to optimize the summary quality but some other peripheral objective. In this work we tackle the global optimization problem using A* search with the training of prediction model intact and demonstrate our method to reduce redundancy within a summary. We use the framework proposed by Aker et al. [1] as a baseline and adapt it to globally improve the summary quality. Our results show significant improvements over the baseline.