<?xml version="1.0" standalone="yes"?>
<Paper uid="A94-1047">
  <Title>FoG: A New Approach to the Synthesis of Weather Forecast Text. In IEEE Expert (Special</Title>
  <Section position="4" start_page="215" end_page="215" type="metho">
    <SectionTitle>
2 Sublanguage Engineering Issues
</SectionTitle>
    <Paragraph position="0"> Early corpus analysis of marine forecasts identified a few kinds of information which were being conveyed with high frequency, as well as a &amp;quot;mixed bag&amp;quot; of phenomena of much lower frequency (e.g., WINDS HIGHER IN FJORDS and FOG LIFTING</Paragraph>
  </Section>
  <Section position="5" start_page="215" end_page="215" type="metho">
    <SectionTitle>
AS WINDS GRADUALLY INCREASE in Arctic regions).
</SectionTitle>
    <Paragraph position="0"> It was decided not to generate those sentence types requiring deep meteorological reasoning, to avoid high implementation cost. Over time, however, there has been some pressure to convey low frequency information which has significant value for marine safety (e.g., unexpectedly high winds in Arctic fjords). Use of a corpus has facilitated bringing low-frequency problems to the attention of system builders and users, so that deliberate design decisions can be made before the system is implemented.</Paragraph>
    <Paragraph position="1"> An early goal in FoG was to generate text with stylistic variation by making use of paraphrase alternatives. Text generation typically provides an opportunity to introduce paraphrase variation, although the traditional problem has been finding ways of choosing from among the possible alternatives (Iordanskaja et al., 1991). However, many instances of apparent free variation turned out to have a tendency toward contextual determination, and it appeared easiest to build these tendencies into strict rules. In other cases individual forecasters voiced a clearcut preference for one variant form, which was subsequently implemented as the unique choice, at least for a given weather centre. The final result was the elimination of paraphrase from the generator. It is not clear that this is optimal, but it has simplified the design and implementation process during a phase when forecasters felt that there were more urgent problems.</Paragraph>
    <Paragraph position="2"> One of the surprises in the development of FoG has been the constant evolution of language usage initiated by forecasters. New phenomena are being introduced (e.g., ultraviolet radiation warnings), other phenomena are de-emphasized, and better ways are found to say the same thing. The reasons for this have been quite varied and often specific to a given forecasting office and its client community. The constant &amp;quot;drift&amp;quot; ofsublanguage usage at individual forecasting sites has led to maintenance of local variant systems. The flow of change requests has confirmed the need to keep the components of FoG in their most declarative and transparent form for easy maintenance.</Paragraph>
    <Paragraph position="3"> Early work showed cases where roughly synonymous words turned out to have somewhat different fuzzy semantics. For example winds can both &amp;quot;strengthen&amp;quot; and &amp;quot;increase&amp;quot;, but the former term tends to be used with high wind speeds. We incorporated a strict separation rule by &amp;quot;legislating&amp;quot; a point on the wind speed scale to separate the two word definitions. In other cases, apparently random variation in usage by forecasters led to an attempt to introduce a reasonable set of criteria for choosing one variant form over another. It appears that the very idea of free variation in forecast wording is difficult for forecasters to accept, and this natural tendency actually makes life easier (but less interesting) for system designers.</Paragraph>
    <Paragraph position="4"> Future extensions to FoG are planned, including new forecast types (e.g., technical synopses) and an option for synthesized speech output, building on the existing linguistic model. We would also like to generate forecasts in languages such as Inuktitut, but this may require a deeper interlingual representation, such as the semantic net already used in other applications (Iordanskaja et al., 1991). However, languages like Inuktitut also use different conceptualizations of the weather than English and French, which might go beyond the capabilities of the FPA.</Paragraph>
    <Paragraph position="5"> Recent attempts to produce spoken forecasts with concatenated speech techniques or commercial text-to-speech output devices suffer from a lack of good prosody. FoG's Meaning-Text language model provides for explicit prosodic structure, percolating from the interlingual representation to a new phonetic representation level. Contrastive stress will come from text planning, while most other features affecting pitch will come from surface syntactic specifications. null</Paragraph>
  </Section>
class="xml-element"></Paper>