<?xml version="1.0" standalone="yes"?>
<Paper uid="W98-1406">
  <Title>De-Constraining Text Generation</Title>
  <Section position="6" start_page="55" end_page="55" type="concl">
    <SectionTitle>
4 Comparison to Other Generation Systems
</SectionTitle>
    <Paragraph position="0"> Related work * exists in two areas: (i) the processing strategy of microplanning tasks, and (ii) the nature and organization of resources used by the microplanner.</Paragraph>
    <Paragraph position="1"> There is a strong tendency in generation to deal with microplanning tasks in a small *number of modules, *which are either structurally or functionally motivated. However, it is recognized that many of the tasks are highly intertwined, so that, in principle, the modules should run in parallel and nearly Constantly exchange information. We consider this as a clear hint that a coarse-grained, task-oriented division of microp!anning sets up artificial barriers. Repeated efforts of researchers to try and breach those barriers confirm our view.</Paragraph>
    <Paragraph position="2"> \[Elhadad et a1.1997\] recognizes that constraints on lexical choice come from a wide variety of sources and are multidirectional, making it difficult to determine a systematic ordering in which they should be taken into account. They propose a backtracking mechanism within a unification framework to *overcome the problem. \[Rubinoff1992\] is perhaps the most strongly focused on this issue. He argues that the accepted division into components &amp;quot;ultimately interferes with some of the decisions necessary in the generation process.&amp;quot; He utilizes annotations as a feedback mechanism to provide the planning stages with linguistically relevant knowledge.</Paragraph>
    <Paragraph position="3"> * Another area of research that belies the unnatural task-based division widely accepted by.text generation researchers today is the attempts to control sentence planning tasks. \[Nirenburg et al. 1989\] and more recently, \[Wanner and Hovy1996\] advocate a blackboard control mechanism, arguing that the order of sentence planning tasks cannot be pre-determined. Behind this difficulty is the real* ity that different linguistic phenomena have different, unpredictable requirements. Grammatical, stylistic and collocati0nal constraints combine at unexpected times during the various tasks of sentence planning.* Blackboard architectures, theoretically, can be used to allow a certain thread of operation to suspend operation until a needed bit of information is available. Unfortunately, in the best case, such an architecture is inefficient and difficult to control. In practice, such systems, as is admitted in both papers above, resort to a *&amp;quot;default (processing) sequence for the modules&amp;quot; along with a simplistic truth-maintenance system which ultimately becomes a fail-and-backtrack type of control, completely negating the spirit of the blackboard* system. While these shortcomings might eventually be *overcome, the fact remains that it was the unnatural division into tasks that necessitated *the blackboard processing in the first place.</Paragraph>
    <Paragraph position="4"> * In this paper, we propose an input data-oriented division of the microplanning task--similar to the way many incremental generators \[De Smedt1990, Reithinger1992, Kilger and Finkler1995\] divide the task of surface processing. However, the processing of input units as done by the Hunter-Gathere r ---our microplanning enginc differs significantly from the processing in the incremental generators cited.* Thus, an important feature of HC is that it possesses a strategy for dividing the problem of verbalizing a semantic structure into relatively independent subproblems. The  subproblems can be of different size. Into which subproblems the problem is divided depends on constraints that hold between units in the input structure. This strategy *greatly contributes to the efficiency of HG. In traditional incremental generators, a unit in the input structure is considered to be a subproblem. Furthermore, HG is bidirectional, i.e., it is usable for both analysis and generation.</Paragraph>
  </Section>
class="xml-element"></Paper>