Headline Generation Based on Statistical Translation

Headline Generation Based on Statistical Translation Michele Banko Computer Science Department Johns Hopkins University Baltimore, MD 21218 banko@cs.jhu.edu Vibhu O. Mittal Just Research 4616 Henry Street Pittsburgh, PA 15213 mittal@justresearch.com Michael J. Witbrock Lycos Inc. 400-2 Totten Pond Road Waltham, MA 023451 mwitbrock@lycos.com Abstract Extractive summarization techniques cannot generate document summaries shorter than a single sentence, some- thing that is often required. An ideal summarization system would under- stand each document and generate an appropriate summary directly from the results of that understanding. A more practical approach to this problem re- sults in the use of an approximation: viewing summarization as a problem analogous to statistical machine trans- lation. The issue then becomes one of generating a target document in a more concise language from a source docu- ment in a more verbose language. This paper presents results on experiments using this approach, in which statisti- cal models of the term selection and term ordering are jointly applied to pro- duce summaries in a style learned from a training corpus. 1 Introduction Generating effective summaries requires the abil- ity to select, evaluate, order and aggregate items of information according to their relevance to a particular subject or for a particular purpose. Most previous work on summarization has fo- cused on extractive summarization: selecting text spans - either complete sentences or paragraphs – from the original document. These extracts are Vibhu Mittal is now at Xerox PARC, 3333 Coyote Hill Road, Palo Alto, CA 94304, USA. e-mail: vmit- tal@parc.xerox.com; Michael Witbrock’s initial work on this system was performed whilst at Just Research. then arranged in a linear order (usually the same order as in the original document) to form a sum- mary document. There are several possible draw- backs to this approach, one of which is the fo- cus of this paper: the inability to generate co- herent summaries shorter than the smallest text- spans being considered – usually a sentence, and sometimes a paragraph. This can be a problem, because in many situations, a short headline style indicative summary is desired. Since, in many cases, the most important information in the doc- ument is scattered across multiple sentences, this is a problem for extractive summarization; worse, sentences ranked best for summary selection of- ten tend to be even longer than the average sen- tence in the document. This paper describes an alternative approach to summarization capable of generating summaries shorter than a sentence, some examples of which are given in Figure 1. It does so by building sta- tistical models for content selection and surface realization. This paper reviews the framework, discusses some of the pros and cons of this ap- proach using examples from our corpus of news wire stories, and presents an initial evaluation. 2 Related Work Most previous work on summarization focused on extractive methods, investigating issues such as cue phrases (Luhn, 1958), positional indi- cators (Edmundson, 1964), lexical occurrence statistics (Mathis et al., 1973), probabilistic mea- sures for token salience (Salton et al., 1997), and the …

Content truncated to 500 words