<?xml version="1.0" standalone="yes"?>
<Paper uid="C02-1113">
  <Title>Natural Language and Inference in a Computer Game</Title>
  <Section position="7" start_page="1" end_page="1" type="concl">
    <SectionTitle>
6 Conclusion and Outlook
</SectionTitle>
    <Paragraph position="0"> We have described an engine for text adventures which uses techniques from computational linguistics to make the interaction with the game more natural. The input is analyzed using a dependency parser and a simple reference resolution module, and the output is produced by a small generation system. Information about the world and about the player's knowledge is represented in description logic knowledge bases, and accessed through a state-of-the-art inference system. Most modules use the inference component; to illustrate its usefulness, we have looked more closely at the resolution and generation of referring expressions, and at the resolution of referential and syntactic ambiguities.</Paragraph>
    <Paragraph position="1"> Preliminary experiments indicate that the performance of our game engine is good enough for fluent gameplay. The constraint based dependency parser we use for parsing and generation achieves very good average case runtimes on the grammars and inputs we use. More interestingly, the inference system also performs very well. With the current knowledge bases, reasoning on the world model and user knowledge takes 546ms per turn on average (with a mean of 39 queries per turn). How well this performance scales to bigger game worlds remains to be seen. One lesson we take from this is that the recent progress in optimizing inference engines for expressive description logics is beginning to make them useful for applications.</Paragraph>
    <Paragraph position="2"> All the language-processing modules in our system are rather simplistic. We can get away with this because the utterances that players seem to want to produce in this setting are restricted, e.g. to objects in the same simulated &amp;quot;location&amp;quot; as the player. (The precise extent of this, of course, remains to be evaluated.) The result is a system which exceeds traditional text adventures by far in the flexibility offered to the user.</Paragraph>
    <Paragraph position="3"> Unlike the input, the output that our game generates is far away from the quality of the commercial text adventures of the eighties, which produced canned texts, sometimes written by professional book authors. A possible solution could be to combine the full generation with a template based approach, to which the TAG-based generation approach we take lends itself well. Another problem is the generation of error messages asking the user to resolve an ambiguous input. The game should ideally generate and present the player with a choice of possible (unambiguous) readings. So, the generation strategy would have to be augmented with some kind of monitoring, such as the one proposed by Neumann and van Noord (1994). Finally, we want to come up with a way of synchronizing the grammars for parsing and generation, in order to ensure that expressions used by the game can always be used by the player as well.</Paragraph>
    <Paragraph position="4"> The system is designed in a way that should make it reasonably easy to replace our simple modules by more sophisticated ones. We will shortly make our adventure engine available over the web, and want to invite colleagues and students to test their own language processing modules within our system. Generally, we believe that the prototype can serve as a starting point for an almost unlimited range of extensions.</Paragraph>
  </Section>
class="xml-element"></Paper>