Generating Extraction-Based Summaries from Hand-Written Summaries by Aligning Text Spans

Human-quality text summarization systems based on sentence extraction are difficult to design because doc- uments can differ along several dimensions, such as length, writing style and lexical usage. The lack of suitable corpor a of extraction-based summaries makes it difficult to evaluate a nd improve existing algorithms. However, there are a large num- ber of hand-written (not extraction-based) summaries avai l- able for news-wire stories. This paper discusses our work on generating a corpus of approximately 25,000 extraction- based summaries from hand-written summaries. We discuss how text-span alignment can be applied to this problem and how this problem differs from previous work on aligning par- allel texts. In addition, we briefly analyze differences bet ween handwritten and extracted summaries.