<?xml version="1.0" standalone="yes"?>
<Paper uid="W97-0706">
  <Title>A Proposal for Task-based Evaluation of Text Summarization Systems</Title>
  <Section position="3" start_page="0" end_page="31" type="intro">
    <SectionTitle>
2 Concepts of Text Summarization
</SectionTitle>
    <Paragraph position="0"> Automatic summaries are usually descnbed m terms of certain key features which relate to the concepts of intent, focus, and Coverage * Intent describes the potential use of the summary, either mdlcattve or reformative Indlcattve summaries, used m this context, provide just enough mformatlon to judge the relevancy of the full text Informattve or substantttve summanes serve as substztutes for the full documents, retaining all tmportant detads * Focus refers to the scope of the summary, either generic or user-directed A generic summary Is based on the mmn concept(s) of a document, whereas a user- or goal-directed summary Is based on the topic of interest indicated by the recipient of the summary * Coverage tn&amp;cates whether the summary is based on a single document or multiple docu&amp;quot; ments Much of the historical work m automatic text summarization has been geared towards the creation of indicative, generic summaries of single documents For example, the work of Luhn (1958), Edmundson (1969), Johnson et al (1993) and Brandow et al (1995) all generated this type of summary, although their approaches have included different combmauons of staustJcal and hngmsttc techmques Luhn (1958) considered frequency of word occurrence within a document  and the posmon of the word m a sentence, Edmundson (1969) looked at cue words, taOe and beading words, and structural indicators, Johnsonet al (1993) used md~tbr phrases, and Brandow et al (1995) apphed sentence welghung using signature word selection Most of these approaches claim some degree of domain independence, however they have been tested only on a specific type of data, such as newspaper arucles (Brandow et al 1995) or techmcal hterature (Edmundson 1969) More recently, the scope of research has expanded to include reformative, user-directed, and multi-document summaries Reamer and Hahn (1988), Maybury (1993), and McKeown and Radev (1995) used knowledge-based approaches to generate mformauve surmnanes that can serve as subsututes for the original document null The expansion m focus to include user-dtrected summaries has been influenced by research m reformation retrieval cormnumty on passage-based retrieval, as m the work of KhanS et al (1996) Also, advances m stalasUcal learning algorithms, such as those maplemented by Kup~ee et al (1995) and Aone et al (1997) have combined generic surnmartes and user-customtzatlon, allowmg the userto affect the content of the summaries by mampulaung sentence extractaon features The potenual for multt-document summarization as proposed by the work of Strzalkowska (1996) and Mare and Bloedom (1997) is based m part on advances m mformauon retrteval and mformauon extracuon performance</Paragraph>
  </Section>
class="xml-element"></Paper>