SWING: Exploiting Category-Specific Information for Guided Summarization

We present our work towards building a robust multiple document summarizer (SWING), with a focus on guided summarization. SWING is an extractive summarizer built upon information retrieval principles. Our key contribution is utilizing category knowledge, collected over all news topics, to calculate category-specific importance (CSI) of sentences. We propose two new category-specific features in this work to compute the CSI of a sentence. We also exploit identified named entities to improve the guided summarization process. Evaluation results show that our methods are effective with respect to both the ROUGE and Pyramid scoring metrics.