<?xml version="1.0" standalone="yes"?>
<Paper uid="C86-1132">
  <Title>VALID UNTIL MIDNIGHT TUESDAY WITH AN OUTLOOK FOR WEDNESDAY. FROBISHER-BAY GALE WARNING ISSUED ... WINDS LIGHT BECOMING SOUTHEASTERLY 15 EAKLY TUESDAY MORNING THEN BACKING AND STRENGTHEN- ING TO EASTERLY 30 TUESDAY AFTERNOON THEN STRENGTHENING TO NORTHEASTERLY GALES 35 TUES- DAY EVENING, MOSTLY CLOUDY WITH SNOW. FOG AND MIST PATCHES. VISIBILITY FAIR IN SNOW, FAIR IN MIST AND POOR IN FOG. OUTLOOK FOR WEDNESDAY GALE FORCE NORTHEASTERLIES BECOMING GALE'- FORCE NORTHER-</Title>
  <Section position="4" start_page="563" end_page="563" type="metho">
    <SectionTitle>
4. A Sample Report
</SectionTitle>
    <Paragraph position="0"> The following simplified example (figure 1) shows the input formatted data, using mnemonic descriptors, for the Frobisher Bay forecast area. (Here, we leave aside the problem of possibly merging reports from the seven areas or sub-areas that are considered together).</Paragraph>
    <Paragraph position="1"> 0400 mon 85/10/16 end.</Paragraph>
    <Paragraph position="2"> frob wind 0 10 &amp; nt 6 dir 140 speed 15 &amp; nt 9 dir 90 speed 30 &amp; nt 6 dir 50 speed 35 &amp; nt 12 dir 20 speed 40  The formatted data identifies the Greenwich time of report validity, the date and area concerned, and then specifies initial values for each important weather parameter. Subsequent changes in the value of a parameter are preceded by the number of hours until the forecast change. Localized exceptions to the general forecast are preceded by a coded sub-area specification. At present, input data is limited to the six most important parameters: (1) wind direction, (2) wind speed, (3) cloud cover classification, (4) precipitation types (if any), (5) precipitation frequency and intensity rating, and (6) air temperature. Further forecast parameters which are functions of the input parameters (e.g., warnings and visibility ratings) are calculated by the first non-linguistic module.</Paragraph>
    <Paragraph position="3"> After reading and analysis, the data is manipulated in clausal form through data checking, area unification and data suppression stages mentioned above. It is then translated into a &amp;quot;logical form&amp;quot; just before input to the linguistic modules.</Paragraph>
    <Paragraph position="4"> Linguistic modules first calculate the values of signficant semantic features of incipient lexical items, particularly regarding direction and degree of changes. For example, winds which change direction in a clockwise direction will be described lexically as &amp;quot;veering&amp;quot; to the new direction, whereas winds which change in a counterclockwise direction are described as &amp;quot;backing&amp;quot;. Initial lexical instantiation uses the most precise term available in the lexicon. Subsequent segmentation into sentences may juxtapose clauses in such a way that lexical variation is desirable. Precise terms may then be replaced by synonymic variants, or by more general (hyperonymic) lexemes.</Paragraph>
    <Paragraph position="5"> Figure 2 gives the final textual form of the marine forecast corresponding to the data of figure 1 above.</Paragraph>
  </Section>
  <Section position="5" start_page="563" end_page="563" type="metho">
    <SectionTitle>
MARINE FORECASTS FOR ARCTIC WATERS ISSUED BY
</SectionTitle>
    <Paragraph position="0"/>
  </Section>
  <Section position="6" start_page="563" end_page="564" type="metho">
    <SectionTitle>
5.Knowledge Sources for Report Synthesis
</SectionTitle>
    <Paragraph position="0"> The RAREAS architecture isolates different types of linguistic and non-linguistic knowledge within appropriate modules. Our grammatical, lexical, rhetorical and stylistic description is based on an examination of all the marine bulletins (manually) produced for the FPCN25 region during the 1983 and 1985 seasons (some 50,000 words in all).</Paragraph>
    <Paragraph position="1"> Examination of this extensive corpus of English has led to a fairly detailed grammar of this sublanguage (cf. Harris 1968, Kittredge and Lehrberger 1982).</Paragraph>
    <Paragraph position="2"> Linguistic knowledge is broken down into several types: - lexical semantics, including conditions for appropriate usage of words in terms of corresponding data values, and frequency preferences among synonymous terms in the subianguage of marine bulletins; - syntactic patterns, including the possible and preferred sentence patterns for expressing messages of given types; a second type of syntactic knowledge concerns the rules for deleting repeated sentence constituents when two or more propositions are fused into a single report sentence; - simple principles of text organization, specific to the variety of text to be synthesized, and hence a function of the data salience hierarchy (see below); Non-linguistic knowledge is of three types: - geographical knowledge for each forecast area including (1) its time zone, (2) its limits of latitude and longitude, and (3) the names of adjoining areas (to allow recursively merging adjacent areas in ease of similar meteorological regimes); - meteorological data including (1) mean temperature values for air and water during each month of the Arctic shipping season (June through October) and (2) record values for temperature &amp; wind speed; - an &amp;quot;archive&amp;quot; of data from preceding reports, used to verify if dangerous wind warnings or freezing spray warnings are in effect.</Paragraph>
    <Paragraph position="3">  Geographic knowledge is used primarily during tire attempt to merge reports for adjoining areas. However time zone data is used to calculate local time associated with meteorological phenomena, and hence allow attribution of appropriate temporal descriptors (e.g., &amp;quot;by late afternoon&amp;quot;). Input data to the system has only the Greenwich reference time used by meteorologists.</Paragraph>
  </Section>
  <Section position="7" start_page="564" end_page="564" type="metho">
    <SectionTitle>
6. Linguistic Treamtent of Salience
</SectionTitle>
    <Paragraph position="0"> The structure of marine weather forecasts shows several linguistic correlates of data salience relations. First, warnings of dangerous conditions (strong winds and freezing spray in the FPCN25 region) constitute separate headers preceding the normal text. Only warnings are so positionally marked and informationally redundant.</Paragraph>
    <Paragraph position="1"> Within the normal text, sentence groups dealing with each forecast parameter are ordered by two principles: inlrinsic interest of the data and implicit causal links between the events or states described. Thus wind direction and speed, as the critical factors in marine conditions, occupy initial position. However visibility ratings, which should tbllow in order of imt)ortance, occur last by virtue of their dependence on fog/mist descriptions, which in turn are somewhat dependent on precipitation, which in aim follow cloud cover ratings. Sentence groups are therefore ordered as follows:</Paragraph>
  </Section>
  <Section position="8" start_page="564" end_page="564" type="metho">
    <SectionTitle>
WINDS &gt; CLOUD-COVER &gt; PRECIP &gt;
FOG&amp;MIST &gt; VISIBILITY
</SectionTitle>
    <Paragraph position="0"> Within each sentence group, sentences and clauses are first ordered according to the dichotomy &amp;quot;general vs. local exception&amp;quot;, and then chronologically within general and exceptional parts. A final correlate of data salience is the choice of marked Iexical items and modifiers. For example, particularly strong winds are classified as &amp;quot;gales&amp;quot; (at 35 knots), &amp;quot;storm force winds&amp;quot; (at 45 knots), etc. Also, more specialized sense verbs such as &amp;quot;veering&amp;quot; and &amp;quot;backing&amp;quot; tend to be used more for huge changes of wind direction.</Paragraph>
    <Paragraph position="1"> 7. Temporal Reference under b~creasing Uncertainty Art interesting problem arises in ascribing particular time adverbials to points and intervals of (local) time. There appears to be a tendency in reportS to &amp;quot;hedge&amp;quot; temporal descriptors slightly as reference time becomes more remote from the forecast issue time. For example, &amp;quot;Tuesday afternomf' or &amp;quot;by (Tuesday) evening&amp;quot; may be preferred for remote reference over the more precise &amp;quot;late Tuesday afternoon&amp;quot;.</Paragraph>
    <Paragraph position="2"> This may reflect the increasing difficulty in predicting onset times for remote meteorological events. RAREAS incorporates two varieties of temporal rules in order to generate more vague temporal descriptors for more remote events.</Paragraph>
    <Paragraph position="3"> 8. Bilingual Reports The RAREAS system was designed to accomodate the synthesis of marine weather bulletins in French as well as in English.</Paragraph>
    <Paragraph position="4"> Only the final three components in the processing sequence are language-dependent (and only the last of these in a non-trivial way).</Paragraph>
    <Paragraph position="5"> Syntactic patterns and lexical enwies for French must of course be furnished on the basis of independent linguistic study of the corresponding French sublanguage. The exact semantics for French (correspondences between data configurations and specific lexemes) must be worked out separately, since there is no guarantee that English and French are lexically one-to-one, even in this narrow domain.</Paragraph>
    <Paragraph position="6"> Canadian weather forecasts of all varieties are currently translated into French by the METEO system (Chevalier et ,-d. 1978), developed at the Universit6 de Montr6al some ten years ago.</Paragraph>
    <Paragraph position="7"> Although METEO takes advantage of the relative closure and stereo.typed style of forecasts, a cerxain percentage of forecast sentences fails analysis and hence translation. This is due not only to input en'ors due to typing and line noise, but also to slight irregularities in the usage of English grammar and lexicon on the part of forecasters, w/erich have proved troublesome to foresee in a compact system.</Paragraph>
    <Paragraph position="8"> The automatic synthesis of marine forecasts, on the other hand, should eliminate the fuzzy edges of unpredictability in truman language production, by using a semantically complete and consistent subset of language to cover all foreseeable data configurations. Work on RAREAS thus prepares the ground ior an attractive alternaave to machine translation of these forecasts. The simultaneous synthesis of English and French forecasts directly from data would optimize the transfer of information to speakers of both languages, in addition to being (in principle) more reliable. Parallel synthesis of bilingual forecasts bypasses translation altogether, and most of the system's work in fact serves for both languages.</Paragraph>
  </Section>
class="xml-element"></Paper>