<?xml version="1.0" standalone="yes"?>
<Paper uid="C86-1126">
  <Title>A GRAMMAR USED FOR PARSING AND GENERATION</Title>
  <Section position="2" start_page="0" end_page="536" type="metho">
    <SectionTitle>
2. THE GRAMMAR FORMALISM
</SectionTitle>
    <Paragraph position="0"> The formalism we have defined allows us to write a single grammar of a language which is used both for analysis and generation by means of two specialized interpreters.</Paragraph>
    <Paragraph position="1"> Sentence analysis is viewed as the transition from a surface structure to a semantic representation which is a Functional Description. Sentence generation is the transformation of a semantic representation into a syntactic form. This symmetry between the two processes has to be clearly expressed if we want a clear notation, easy to read and to understand from the point of view of parsing and of generating.</Paragraph>
    <Paragraph position="2"> A grammar rule is itself represented as a Functional Description. This FD has three main &amp;quot;identifiers&amp;quot; : PATTERN, MEANING and CONSTRAINTS.</Paragraph>
    <Paragraph position="3">  Example of a simple grammar rule :</Paragraph>
    <Paragraph position="5"> The ~ part describes the syntactical structure. Each item of the list associated to pattern refers to a rule or to a terminal. In the above example the three terms refer to terminals. Omissions and repetitions are allowed.</Paragraph>
    <Paragraph position="6"> The meaninq part describes the semantic representation associated to the syntactical structure. Bracketed lists represent &amp;quot;paths&amp;quot; refering to Functional Descriptions inside the rule or in another rule. During parsing, these paths are used to build the semantic representation while in generation they are used for splitting a semantic structure into different sub-structures. The two processes, parsing and generation, are detailed in chapters 3 and 4.</Paragraph>
    <Paragraph position="7"> The constraints part is a list of &amp;quot;set of constraints&amp;quot; expressed by Functional Descriptions. At \]east one &amp;quot;set of constraints&amp;quot; must be fulfilled. In the above example this allows us to express agreement rules used for both parsing and generating.</Paragraph>
    <Paragraph position="8"> As in Martin Kay definitions a rule may have different derivations. These are represented by enclosed braces. Example :</Paragraph>
    <Paragraph position="10"/>
  </Section>
  <Section position="3" start_page="536" end_page="536" type="metho">
    <SectionTitle>
3. THE PARSING PROCESS
</SectionTitle>
    <Paragraph position="0"> 3.1. Us___e of th e ~rammar for In order to anaiyze a sentence, the words and compounds words are converted in Functional Descriptions, using a morphological analyzer and a dictionary. The result is a list of FD's which will be processed by -the parser.</Paragraph>
    <Paragraph position="1"> Example (semantic va\]ues are expressed here by English terms but they are usually expressed as FD) : &amp;quot;\]es chaussures vertes&amp;quot; (&amp;quot;the green shoes&amp;quot;) Input list of parser is : ( \[cat = det \[cat = subst \[cat = adj type =defined gender = fem gender = fem number:plural number:plural number=plural lex : \]e\] lex=ehaussure lex = vert meaning=shoe\] meaning:green\] This sentence matches with the rule simple_gn described in chapter 2, as the first FD of the list is funchionnaly unifiab\]e with \[cat = de t\], the second FD with \[cat = subst\] and the third FD with \[cat = adj\].</Paragraph>
    <Paragraph position="2"> The parsing process builds a structure which is a copy of the rule simple_gn and enlarges it with the actual, word analyzed. The path descriptions are replaced by their actual values.</Paragraph>
    <Section position="1" start_page="536" end_page="536" type="sub_section">
      <SectionTitle>
3.2. Structure built
</SectionTitle>
      <Paragraph position="0"/>
      <Paragraph position="2"> This structure is built if the constraints are met : for this rule it implies agreement of gender and number, which is the case for &amp;quot;les chaussures vertes&amp;quot;.</Paragraph>
    </Section>
  </Section>
  <Section position="4" start_page="536" end_page="538" type="metho">
    <SectionTitle>
4. THE GENERATING PROCESS
</SectionTitle>
    <Paragraph position="0"> 4.1o Use of the ~rammar :for ~eneration The generation-takes as input a semantic structure and produces a sentence.</Paragraph>
    <Paragraph position="1"> As an example the rule simple_gn (cf chapter 2), is activated with the semantic structure</Paragraph>
    <Paragraph position="3"> A copy of the rule is built. The paths in * the Functional Description associated to the identifier &amp;quot;meaning&amp;quot; are used to convey the semantic information to &amp;quot;the items referred to by the identifier &amp;quot;pattern&amp;quot; (These items are named &amp;quot;constituents&amp;quot;) \]~n this example Jt gives :  The interpretation process of the grammar &amp;quot;builds&amp;quot; the path, which means that the needed identifiers are included in the copy of -the rule.</Paragraph>
    <Paragraph position="5"> where &amp;quot;meaning&amp;quot; and &amp;quot;number&amp;quot; have been</Paragraph>
    <Paragraph position="7"> where &amp;quot;meaning&amp;quot; has been added.</Paragraph>
    <Paragraph position="8"> Then the constraints are applied. In the parsing process they are used to eliminate wrong constructions, while in the generating process they are used to transmit information.</Paragraph>
    <Paragraph position="9"> Use of the constraints of the rule simple gn</Paragraph>
    <Paragraph position="11"> At this step, this rule doesn't transmit any information because identifier &amp;quot;gender&amp;quot; is not present in at least one Functional</Paragraph>
    <Paragraph position="13"> This rule transmits number of substantive (number = plural), in the two other Functional descriptions of the output list After constraints are applied, the output</Paragraph>
    <Paragraph position="15"> The next step is word selection : for each terminal, the semantic structure associated with it is used to choose a lexical item.</Paragraph>
    <Paragraph position="16"> This is done by using Functional Unification. For each word or compound word selected, &amp;quot;constraints&amp;quot; are processed again, in order to transmit informations to Functional Descriptions of the list.</Paragraph>
    <Paragraph position="17"> For a given structure there may be more than one adequate word. In that case the appropriate word is chosen by the user interactively.</Paragraph>
    <Paragraph position="18"> The list of terminals is enlarged by the selected lexical items, as shown in the following example : For the first item : (  At this step each word of the output list is completely defined. The morphological generation processes each Functional Description using fields LEX, GENDER, NUMBER, MODE, TENSE and PERSON. The appropriate form of the lexieal item is constructed using Functional Unification. For this example the list constructed by the morphological generation is : ( &amp;quot;des&amp;quot;, &amp;quot;boites&amp;quot;, &amp;quot;blanches&amp;quot; ) which gives : &amp;quot;des boites blanches&amp;quot; This example is a simple case where items of a &amp;quot;pattern&amp;quot; do not refer to other rules. Presence of a rule name in a pattern leads to activation of this rule with a subset of the initial meaning (transmitted by a path, as for a terminal).</Paragraph>
    <Paragraph position="19"> 4.2. Generation models The generation of the sentence associated to a semantic structure may lead to various syntactical constructs. In order to reduce the number of constructs, and to allow control of text style, a specific feature has been introduced, named &amp;quot;generation model&amp;quot;. A generation model associates a semantic pattern to a precise grammar rule. Example : Semantic structure associated to the advice &amp;quot;Do not expose to rain&amp;quot; in a user's manual :</Paragraph>
    <Paragraph position="21"> Among the &amp;quot;generation models&amp;quot; of the system, the following is Functionnaly Unifiable to the above structure :</Paragraph>
    <Paragraph position="23"> Remark : the symbol * means that the rule may be repeated.</Paragraph>
    <Paragraph position="24"> This generation model is selected by a restricted version of Functional Unification : identifiers &amp;quot;advice&amp;quot; and &amp;quot;advice-type&amp;quot; must be present in the semantic structure.</Paragraph>
    <Paragraph position="25"> In this example two grammar rules are candidate once the generation model is selected. A simple implementation is to choose a rule at random, another is to have an evaluation module which choose the most appropriate rule according to stylistic knowledge (technical style, telegraphic style, etc).</Paragraph>
  </Section>
  <Section position="5" start_page="538" end_page="538" type="metho">
    <SectionTitle>
5. DEVELOPMENTS
</SectionTitle>
    <Paragraph position="0"> Previous version of the multilingual generation system uses a grammar for parsing, and production rules for generation.</Paragraph>
    <Paragraph position="1"> Present work is the adaptation of the parser to the new formalism, and the implementation of the generation interpreter. It includes the adaptation of the multilingual dictionary retrieval process.</Paragraph>
  </Section>
class="xml-element"></Paper>