Framework for Abstractive Summarization using Text-to-Text Generation

We propose a new, ambitious framework for abstractive summarization, which aims at selecting the content of a summary not from sentences, but from an abstract representation of the source documents. This abstract representation relies on the concept of Information Items (INIT), which we define as the smallest element of coherent information in a text or a sentence. Our framework differs from previous abstractive summarization models in requiring a semantic analysis of the text. We present a first attempt made at developing a system from this framework, along with evaluation results for it from TAC 2010. We also present related work, both from within and outside of the automatic summarization domain.