<?xml version="1.0" standalone="yes"?>
<Paper uid="P84-1020">
  <Title>LIMITED DOMAIN SYSTEMS FOR LANGUAGE TEACHING</Title>
  <Section position="5" start_page="85" end_page="86" type="concl">
    <SectionTitle>
DEVELOPMENTS
</SectionTitle>
    <Paragraph position="0"> The mechanism is currently embedded within two small domains. The one illustrated here is 'told' a simple 'story' and then asks or answers questions about that. The sample grammar was intended to demonstrate the interaction of wh questions with passives, among other things.</Paragraph>
    <Paragraph position="1"> Although we are not here concerned with the semantics of these domains, they are fairly simple, and several different types of semantic components are used depending on the nature of the domain. For some domains a procedural semantics is appropriate, manipulating objects on  a screen or asking and answering questions about them. In the 'William' program here a production system again based on the Pop-ll matching procedures is used, currently being coupled to a simple backwards chaining inference mechanism.</Paragraph>
    <Paragraph position="2"> Neither the grammatical routines nor any embodiment of them constitute a complete tuition system, or anything approaching that: they are merely frameworks for experimentation. But the syntactic error detection routines could be used in many other environments where useful feedback of this type was required, say in database interrogation or machine translation. Within a language tuition context the mechanism could be used to advantage without an associated semantics, in some of the more traditional types of computer aided EFL teaching programs: for example, gap-filling, drill and practice, sentence completion, or grammatical paraphrase tasks. Only trivial adjustments would be needed to the overall mechanism for this to become a powerful and sophisticated framework within which to elaborate such programs.</Paragraph>
    <Paragraph position="3"> However, there are several ways in which the general mechanism might be improved upon, most immediately, the following: (i) if a parse fails early in the sentence, the user only gets a report based on that part of the sentence, when there may be more serious errors later one (or some praiseworthy use of the language). In these cases a secondary parse looking for well-formed sub-constituents, in something like the way a chart parser might do, would provide useful information. (I am grateful to Steve Isard and Henry Thompson for this suggestion).</Paragraph>
    <Paragraph position="4"> (ii) the quality of the example continuations could be improved. Eventually it would be desirable to have the generator semantically guided, but this is by no means trivial, even in a limited domain. There are several heuristics which can produce a better type of continuation, however: using a temporary lexicon containing words from the unparsed portion of the sentence, or from the most recently parsed sentences, or combinations of these with the restricted sub-lexicon. In the best cases this type of heuristic can be spectacularly successful, producing a grammatical version of what the user was trying to say. However, they can also flop badly: more testing on real students would be one way of disceovering which of these alternatives is best.</Paragraph>
    <Paragraph position="5"> (iii) as suggested in Weischedel and Black, it might be profitable to explore the use of semantic grammars - grammars using semantically rather than syntactically motivated categories in the system. Although of dubious theoretical status, they are a useful engineering tool: the non-terminals can be labelled in a domain-specific way that is transparent for the user, and, being semantically motivated, the system could appear as if it were doing semantic diagnosis of a limited type as well as syntactic diagnosis. For example, instead of being prompted for an adjective, the user might be prompted for 'a word describing the appearance of a car', or something equally specific. Furthermore, the availability of the pre-compilation programs means that it should be possible to use the metarule formalism for these grammars also: this should go some way towards minimising their linguistic disadvantages, namely, a tendency to repetition and redundancy in expressing facts about the languages they generate.</Paragraph>
    <Paragraph position="6"> The system is written in Pop-ll (a Lisp-like language) within the POPLOG programming environment developed by the University of Sussex. At UEA POPLOG runs on a VAX 11/780 under VMS.</Paragraph>
  </Section>
class="xml-element"></Paper>