<?xml version="1.0" standalone="yes"?>
<Paper uid="E93-1055">
  <Title>Lexical Choice Criteria in Language Generation</Title>
  <Section position="7" start_page="456" end_page="457" type="concl">
    <SectionTitle>
6 Summary and Future Work
</SectionTitle>
    <Paragraph position="0"> An important task in language generation is to choose the words that most adequately fit into the utterance situation and serve to express the intentions of the speaker. I have listed a number of criteria for lexical choice and then explored stylistic dimensions in more detail: Arguing in favour of a 'data-driven' approach, sets of synonyms have been extracted from thesauri and dictionaries; comparing them led to a proposed set of features that can discriminate synonyms on stylistic grounds. The features chosen in the implementation have been selected solely on the basis of the author's intuitions (albeit using a systematic method) -- clearly, these findings have to be validated through psychological experiments (asking subjects to compare words and rate them on appropriate scales). Also, it needs to be explored in more detail whether different parts of speech should be  characterized by different feature sets.</Paragraph>
    <Paragraph position="1"> An overall model of lexicalization in the generation process has been sketched that first determines all candidate lexical items for expressing parts of a message (including all synonyms and less-specific items), and a preferential choice process is supposed to make the selections. The front-end of this system has been implemented by extending the PENMAN sentence generator so that it can choose words on the basis of a distance function that compares the feature/value pairs of lexical entries (of synonyms) with a target specification. This target specification has so far only been postulated as corresponding to various stereotypical genres, the name of which is a part of the input specification to PENMAN. In future work, the stylistic features need to be linked more systematically to rhetorical goals of the speaker and to parameters characterizing the utterance situation. One of the tasks here is to determine which features should be valid for the whole text to be generated (e.g., formality), or only for single sentences, or only for single constituents (e.g., slant).</Paragraph>
    <Paragraph position="2"> Besides, ultimately the work on lexical style has to be integrated with efforts on syntactic style \[DiMarco and Hirst, 1993\]. Other criteria for lexical choice, like those mentioned in section two, have to be incorporated into the choice process. And finally, it has to be examined how lexical decisions interact with the other decisions to be made in the generation process (sentence scope and structure, theme control, use of conjunctions and cue words, etc.).</Paragraph>
  </Section>
class="xml-element"></Paper>