<?xml version="1.0" standalone="yes"?>
<Paper uid="J89-4002">
  <Title>NATURAL LANGUAGE GENERATION FROM PLANS</Title>
  <Section position="26" start_page="248" end_page="248" type="concl">
    <SectionTitle>
5.4 WAYS FORWARD
</SectionTitle>
    <Paragraph position="0"> In this paper we have described a system for generating natural language from automatically generated plans.</Paragraph>
    <Paragraph position="1"> Our main aim in developing the system was to produce a model of a complete system using state-of-the-art methodology and techniques, partly to evaluate the current state of knowledge, and partly to provide a basis for comparison for future work. Logically, then, there are two strands to further work based on this research: building on the evaluation to learn lessons about the design of generation systems and the systems they interact with, and building on the system itself to produce better generation-from-plans systems.</Paragraph>
    <Paragraph position="2"> One of the key evaluative lessons concerns the plan structures a system like this depends on. We found the plan,; produced by NONLIN unsatisfactory and we have begun to understand why. We must now specify what we would like a plan to look like and contain, within the general constraint that a planning system couht reasonably produce it. Our present approach to Comlmtational Linguistics Volume 15, Number 4, December 1989 Chris Meilish and Roger Evans Natural Language Generation from Plans this topic is to take a very formal view of plans as algebraic expressions over states (rather than actions or goals) with a well-defined formal semantics, allowing us to be clear about the semantic effect of plan transformations. null The system itself falls broadly into two parts, building and simplifying the message, and turning the message into text. Of these the latter is the more modular, more declarative, and probably more successful at present. To a certain extent it can serve as a piece of enabling technology for research on the message component. Its major deficiency as discussed above is global stylistic control. Its handling of morphology is currently rather unprincipled, but the utilization of a morphological representation language such as Datr (Evans and Gazdar 1989 a,b) would rectify this.</Paragraph>
    <Paragraph position="3"> The biggest outstanding task, however, is the message planner itself. The mechanism described above employs some quite powerful techniques in a fairly effective way, but it is not very perspicuous or extensible. We have begun work on a new message planner module that applies transformation rules to plans of the algebraic type mentioned above, gradually transforming the plan into an optimized message structure. This will provide us with a rule-based semideclarative framework in which to explore further the issues of message planning.</Paragraph>
  </Section>
class="xml-element"></Paper>