A Web-Trained Extraction Summarization System

A Web-Trained Extraction Summarization System Liang Zhou and Eduard Hovy USC Information Sciences Institute 4676 Admiralty Way Marina del Rey, CA 90292-6695 {liangz, hovy}@isi.edu Abstract A serious bottleneck in the development of trainable text summarization systems is the shortage of training data. Constructing such data is a very tedious task, especially because there are in general many different correct ways to summarize a text. Fortunately we can utilize the Internet as a source of suitable training data. In this paper, we present a summarization system that uses the web as the source of training data. The procedure involves structuring the articles downloaded from various websites, building adequate corpora of (summary, text) and (extract, text) pairs, training on positive and negative data, and automatically learning to perform the task of extraction-based summarization at a level comparable to the best DUC systems. 1 Introduction The task of an extraction-based text summarizer is to select from a text the most important sentences that are in size a small percentage of the original text yet still as informative as the full text (Kupiec et al., 1995). Typically, trainable summarization systems characterize each sentence according to a set of predefined features and then learn from training material which feature combinations are indicative of good extract sentences. In order to learn the characteristics of indicative summarizing sentences, a large enough collection of (summary, text) pairs must be provided to the system. Research in automated text summarization is constantly troubled by the difficulty of finding or constructing large collections of (extract, text) pairs. Usually, (abstract, text) pairs are available and can be easily obtained (though not in sufficient quantity to support fully automated learning for large domains). But abstract sentences are not identical to summary sentences and hence make direct comparison difficult. Therefore, some algorithms have been introduced to generate (extract, text) pairs expanded from (abstract, text) inputs (Marcu, 1999). The explosion of the World Wide Web has made accessible billions of documents and newspaper articles. If one could automatically find short forms of longer documents, one could build large training sets over time. However, one cannot today retrieve short and long texts on the same topic directly. News published on the Internet is an exception. Although it is not ideally organized, the topic orientation and temporal nature of news makes it possible to impose an organization and thereby obtain a training corpus on the same topic. We hypothesize that weekly articles are sophisticated summaries of daily ones, and monthly articles are summaries of weekly ones, as shown in Figure 1. Under this hypothesis, how accurate an extract summarizer can one train? In this paper we first describe the corpus reorganization, then in Section 3 the training data formulation and the system, the system evaluation in Section 4, and finally future work in Section 5. daily weekly monthly Figure 1. Corpus structure. Edmonton, May-June 2003 Main Papers , pp. 205-211 Proceedings of HLT-NAACL 2003 2 Corpus Construction 2.1 Download Initial Collection The Yahoo Full Coverage Collection (YFCC) …

Content truncated to 500 words