<?xml version="1.0" standalone="yes"?>
<Paper uid="P98-2149">
  <Title>Dealing with distinguishing descriptions in a guided composition system</Title>
  <Section position="2" start_page="908" end_page="910" type="abstr">
    <SectionTitle>
1 The project EREL is partially funded by the
</SectionTitle>
    <Paragraph position="0"> Conseil Gtntral des Bouches-du-RhSne.</Paragraph>
    <Paragraph position="1"> educational play activities designed to help users to employ common language.</Paragraph>
    <Paragraph position="2"> So, it is very important in this kind of applications to offer a sophisticated guided composition mode, and in particular to produce only correct sentences at every level (syntactical, conceptual but also contextual). At the moment, there are two main types of games in EREL: in the first one, chidren speak about or ask questions about a picture they see on the screen. The context which contains objects about which the chidren can tall about is preliminary computed and it does not change along with the discourse.</Paragraph>
    <Paragraph position="3"> So, the set of distinguishable entities DE is built at one time.</Paragraph>
    <Paragraph position="4"> The second activity concerns a dialog on a logic game in which users compose orders in natural language to achieve a goal. One of the exercises consists in putting and moving objects on a board. A child has an initial stock of objects that he can put on a checker board, permute, move, or stow away. He gives orders to the system using natural languages sentences and he can see immediately on the board the effects the sentences have. The interface looks like this:  Here, in the French version, the user has begun a sentence ~,Echange le carr~ noir avec le rond...~ (Permute the black square with the circle...) and the system, according to the contextual situation, proposes the possible words to be selected: blanc, gris, noir (white, grey, black). If there had been no white circle on the board, the word white would not have been proposed.</Paragraph>
    <Paragraph position="5"> We have also implemented part of the -inclusion presented above so the child can use the hyponym pawn to designate a triangle or a square or a circle. The hyponym relations are already represented in the  system by a conceptual graph which is used in other parts of the system.</Paragraph>
    <Paragraph position="6"> In this game, it is clear that contextually correct definite descriptions must designate objects in the context (the system has to act on them, so it has to find them). So, in the guided composition mode, the system has to compute/&amp;/'as they have been defined in SS2. The objects are moving during the game so, as opposed to the ira'st type of activity, the context changes and the set of distinguishable entities has to be computed at every step. Moreover, the children can create new pawns (made from a predefined set of shapes and colors). These objects are added in the context and must be taken into account for later mentions.</Paragraph>
    <Paragraph position="7"> The set of distinguishable entities is not the set of all the objects because some objects may not be distinguishable. For example, if two objects have the same shape and color and are in the stock ('Rtserve' on the figure above), then they can not be distinguished (we assume there is no relative position in the reserve). The user can act on them by using sentences like 'put one of the red triangles which are in the reserve in the case A4' but not like 'put the red triangle which is in the reserve...'. If there is no other triangle on the chessboard, then the system must not propose beginnings of definite NP like 'the triangle...' EREL is under development and a medical team who works with autistic children is testing a preliminary version.</Paragraph>
    <Paragraph position="8"> 4.Related works The work presented here uses notions firstly introduced by \[Dale 89\] and mentioned in many works in the field. We have presented here new applications and extensions. Actually, the problem treated here raises the general question of generating definite descriptions. Generally, these works deal rather with the problem of what to say and how to say it. \[Novak 88\] deals with the problem (among others) of when and how restrictives relative clause have to be used in definite NP. The system that he describes is able to produced definite NP like the first yellow BMW, the second yellow BMW if necessary. \[Kronfeld 89\] talks about 'conversionally relevant descriptions' which is typically the problem of how to say something according to the context of discourse or the user's goal like in \[Appelt 85\]. The relations between Gricean maxims and the generation of definite NP are studied in \[Passonneau 95\] and \[Dale &amp; al. 96\] for example.</Paragraph>
    <Paragraph position="9"> \[Horacek 97\] gives a good comparison of the previous works; his analyses make appear the problem of the linguistic realisation of a set of properties of an entity to generate a description that designates it and he proposes an algorithm which takes into account this problem during the choice of the property which will be used to build a Concerning the production of Idd; we are not really confronted with the problems mentioned in \[Horacek 97\] because the guided composition system doesn't generate NPs from entity representation; its parser generates partial syntactically correct sentences which are filtered by contextual criteria (the processes are driven in parallel thanks to coroutined methods). Moreover, concerning what to say and how to say it, it is the user who chooses what word (among the possibilities offered by the system) will be kept to build the sentence, at every step. Concerning the extension of the notion of agreement that we make (and so of the notion of distinguishing description), many linguists mention the phenomenon we want to take into account. A more computational point of view is discussed in \[Groenendijk &amp; al 96, pp 25-27\] (the example of 'the doctor' and 'the man'). The authors do not give really computable solutions to this problem. It seems that the use of simple inclusion and ~inclusion to f'md distinguishable entity and to identify a referent for a (complete or incomplete) distinguishing description (as described in SS2.2) deals rather efficiently with the problem 5.Conclusion We showed here how the uniqueness requirement, when dealing with incomplete definite descriptions, turns into a requirement of that particular sort of entities from the  context, the distinguishable entities. Then we showed how the notion of distinguishing description can be extended using inclusion and what we called ~inclusion. An algorithm that uses these ideas and allows to know as early as possible incomplete definite description that earl lead to correct definite description from those that cannot is given. The algorithm is incremental, which is partieulary useful in a guided composition system and allows also to solve complete definite description (finding the referent). So far, an instance of it has been implemented under the system EREL.</Paragraph>
  </Section>
class="xml-element"></Paper>