Language Modeling Experiments in Non-Extractive Summarization

Although most text summarization research to date has been applied to news articles, web pages are quite different in both structure and content. Instead of coherent text with a well-defined discourse structure, they are mostly a bag of phrases, links, graphics and formatting commands, thus providing few opportunities for extractive summarization methods. Extractive summarizers, moreover, are limited in their ability to produce very brief, headline-like, summaries where flexibility in lexical choice and phrasing are important. This paper discusses relatively simple statistical models for generating non-extractive summaries of web pages. It describes the datasets used to train these models, shows sample outputs, and discusses the results of some preliminary evaluations to assess the quality of the resulting summaries.