<?xml version="1.0" standalone="yes"?>
<Paper uid="E03-1062">
  <Title>A Flexible Pragmatics-driven Language Generator for Animated Agents</Title>
  <Section position="8" start_page="0" end_page="0" type="concl">
    <SectionTitle>
4 Implementation
</SectionTitle>
    <Paragraph position="0"> The NECA MNLG has been implemented in PRO-LOG. The output is in the form of an RRL XML document. Table 1 provides a sample of the response times of the compiled code running on a Pentium III Mobile 1200 Mhz with Sicstus 3.8.5 PROLOG. We timed the complete generation process from parsing the XML input to producing XML output, including generation of deep syntactic structure, referring expressions, turn taking gestures (not discussed in this paper), etc.</Paragraph>
    <Paragraph position="1"> input # acts = 1 * 10  The results show generation times for entire dialogues and according to whether the generator was asked to produce exactly one solution or select at random a solution from a set of at most ten generated solutions (the latter strategy was implemented to obtain more variation in the generator output). On average for = 1 the generation time for an individual dialogue act is almost 1100 of a second. For * 10 it is 4100 of a second. The generator uses a repository of 138 trees (including the two examples given above). The repository has been developed for and integrated into the ESHOWROOM system which is currently being fielded. A start is being made with porting the MNLG to a new domain and documentation is being created to allow our project partners to carry out this task. We hope that our efforts will contribute to addressing a challenge expressed in (Retem, supporting fast generation. Moreover, by using features for unbounded dependencies we do not require the adjunction operation, which is incompatible with our topdown generation approach. We follow Nicolov et al. (1996), who also use TAG, in their commitment to flat semantics. Their generator does, however, not take pragmatic constraints into account. iter, 1999): &amp;quot;We hope that future systems such as STOP will be able to make more use of deep techniques, because of advances in linguistics and the development of reusable wide-coverage NLG components that are robust, well-documented and well engineered as software artifacts.&amp;quot; In our view the best way to approach this goal is by providing a framework which allows for the flexible integration of shallow and deep generation, thus making it possible that in the course of various projects, deep analyses can be developed alongside the shallow solutions which are difficult to avoid altogether in software development projects, due to the pressure to deliver a complete system within a certain span of time.</Paragraph>
  </Section>
class="xml-element"></Paper>