<?xml version="1.0" standalone="yes"?>
<Paper uid="W97-0713">
  <Title>From discourse structures to text summaries</Title>
  <Section position="5" start_page="86" end_page="87" type="metho">
    <SectionTitle>
4 Comparison with other work
</SectionTitle>
    <Paragraph position="0"> We are not aware of any RST-based summanzatlon program for Enghsh However, Ono et al (1994) discuss a summanzaUon program for Japanese whose m~mmal textual umts are sentences Due to the differences between Enghsh and Japanese, R was impossible for us to compare Ono's summarizer wtth ours Fundamental differences concerning the assumpttons that underhe Ono's workand ours are discussed at length m (Mareu, 1997b)  An evaluauon of our summarization program We were able to obtmn only one other program that summarizes Enghsh text m the one included m the Macrosoft Office97 package We run the Microsoft summanzaUon program on the five texts from Sczent~fic Amerscan and selected the same percentages of textual umts as those considered Important by the judges When we selected percentages of text that corresponded only to the clauses considered important by the judges, the lVherosoft program recalled 28% of the umts, with a prec~slon of 26% When we selected percentages of text that corresponded to Sentences considered lmportsnt by thejudgus, the Microsoft program recalled 41% of the units, wxth a precision of 39% All Microsoft figures are only shghtly above those that correspond to the basehne algorithms that select Hnportant umts randomly It follows that our program outperforms slgmficantly the one found m the Office97 package We are not aware of any other summanzatton program that can bmld summaries with granularity as fine as a clause (as our program can)</Paragraph>
  </Section>
class="xml-element"></Paper>