Text Generation for Abstractive Summarization

We have begun work on a framework for abstractive summarization and decided to focus on a module for text generation. For TAC 2010, we thus move away from sentence extraction. Each sentence in the summary we generate is based on a document sentence but it usually contains a smaller amount of information and uses fewer words. The system uses the output of a syntactic parser for a sentence and then regenerates part of the sentence using a Natural Language Generation engine. The sentences of the summary are selected among regenerated sentences based on the document frequency of contained words, while avoiding redundancy. Date and location were handled and generated especially for cluster categories 1 and 2. Even though our initial scores were not outstanding, we intend to continue work on this approach in the coming years.