<?xml version="1.0" standalone="yes"?>
<Paper uid="P00-1041">
  <Title>Headline Generation Based on Statistical Translation</Title>
  <Section position="6" start_page="0" end_page="0" type="concl">
    <SectionTitle>
5 Conclusions and Future Work
</SectionTitle>
    <Paragraph position="0"> This paper has presented an alternative to extractive summarization: an approach that makes it possible to generate coherent summaries that are shorter than a single sentence and that attempt to conform to a particular style. Our approach applies statistical models of the term selection and term ordering processes to produce short summaries, shorter than those reported previously. Furthermore, with a slight generalization of the system described here, the summaries need not contain any of the words in the original document, unlike previous statistical summarization systems. Given good training corpora, this approach can also be used to generate headlines from a variety of formats: in one case, we experimented with corpora that contained Japanese documents and English headlines. This resulted in a working system that could simultaneously translate and summarize Japanese documents.8 The performance of the system could be improved by improving either content selection or linearization. This can be through the use of more sophisticated models, such as additional language models that take into account the signed distance between words in the original story to condition 8Since our initial corpus was constructed by running a simple lexical translation system over Japanese headlines, the results were poor, but we have high hopes that usable summaries may be produced by training over larger corpora. the probability that they should appear separated by some distance in the headline.</Paragraph>
    <Paragraph position="1"> Recently, we have extended the model to generate multi-sentential summaries as well: for instance, given an initial sentence such as &amp;quot;Clinton to meet visit MidEast.&amp;quot; and words that are related to nouns (&amp;quot;Clinton&amp;quot; and &amp;quot;mideast&amp;quot;) in the first sentence, the system biases the content selection model to select other nouns that have high mutual information with these nouns. In the example sentence, this generated the subsequent sentence &amp;quot;US urges Israel plan.&amp;quot; This model currently has several problems that we are attempting to address: for instance, the fact that the words co-occur in adjacent sentences in the training set is not sufficient to build coherent adjacent sentences (problems with pronominal references, cue phrases, sequence, etc. abound). Furthermore, our initial experiments have suffered from a lack of good training and testing corpora; few of the news stories we have in our corpora contain multi-sentential headlines.</Paragraph>
    <Paragraph position="2"> While the results so far can only be seen as indicative, this breed of non-extractive summarization holds a great deal of promise, both because of its potential to integrate many types of information about source documents and intended summaries, and because of its potential to produce very brief coherent summaries. We expect to improve both the quality and scope of the summaries produced in future work.</Paragraph>
  </Section>
class="xml-element"></Paper>