<?xml version="1.0" standalone="yes"?>
<Paper uid="C02-1113">
  <Title>Natural Language and Inference in a Computer Game</Title>
  <Section position="3" start_page="0" end_page="0" type="intro">
    <SectionTitle>
2 Architecture
</SectionTitle>
    <Paragraph position="0"> The general architecture of the game engine, shown in Fig. 2, consists of a number of language-processing modules (drawn as ellipses), which in-Observation Lounge This is where the station staff and visitors come to relax. There are a lot of tables and chairs here, a large observation window, and a plush carpet. In the corner you can see an AstroCola dispenser. A tube leads up to the station's main corridor. &gt; put my galakmid coin into the dispenser Click.</Paragraph>
    <Paragraph position="1"> The dispenser display now reads &amp;quot;Credit = 1.00&amp;quot;. &gt; push diet astrocola button You hear a rumbling noise in the dispenser, but nothing appears in the tray. &gt; kick dispenser A can drops into the tray. Amazing! The oldest trick in the book, and it actually worked.</Paragraph>
    <Paragraph position="2">  (drawn as rectangles). There are two separate knowledge bases, which share a set of common definitions: One represents the true state of the world in a world model, the other keeps track of what the player knows about the world. Solid arrows indicate the general flow of information, dashed arrows indicate access to the knowledge bases.</Paragraph>
    <Paragraph position="3"> The user's input is first parsed using an efficient parser for dependency grammar (Duchier and Debusmann, 2001). Next, referring expressions are resolved to individuals in the game world. The result is a ground term or a sequence of ground terms that indicates the action(s) the user wants to take. The Actions module looks up these actions in a database (where they are specified in a STRIPS-like format), checks whether the action's preconditions are met in the world, and, if yes, updates the world state with the effects of the action.</Paragraph>
    <Paragraph position="4"> The action can also specify effects on the user's knowledge. This information is further enriched by the Content Determination module; for example, this module computes detailed descriptions of objects the player wants to look at. The Reference Generation module translates the internal names of individuals into descriptions that can be verbalized. In the last step, an efficient realization module (Koller and Striegnitz, 2002) builds the output sentences according to a TAG grammar. The player knowledge is updated after Reference Generation when the content of the game's response, including the new information carried e.g. by indefinite NPs, is fully established.</Paragraph>
    <Paragraph position="5"> If an error occurs at any stage, e.g. because a pre-condition of the action fails, an error message specifying the reasons for the failure is generated by using the normal generation track (Content Determination, Reference Generation, Realization) of the game.</Paragraph>
    <Paragraph position="6"> The system is implemented in the programming language Mozart (Mozart Consortium, 1999) and provides an interface to the DL reasoning system RACER (Haarslev and M&amp;quot;oller, 2001), which is used for mainting and accessing the knowledge bases.</Paragraph>
  </Section>
class="xml-element"></Paper>