One of the worst problems with classical text adventures is the "identification problem", which occurs when the game cannot parse its own output. This happens because the output is canned text, and the parser is simple -- but it is very annoying for the user when they can't use e.g. a referring expression that the game just used.
In the current implementation of the game, this problem isn't really solved because it uses different grammars for parsing and realization - -- a dependency grammar for parsing, and a TAG grammar for realization.
This project is about working out how to do realization purely with XDG, the dependency grammar formalism on which our parser is based. This has the potential to solve the identification problem completely if the same grammar can be used for parsing and generation. It is also independently interesting because nobody has worked on XDG realization before.
XDG allows the grammar writer to specify different levels of representation and link them with declarative constraints. We are thinking about representing semantics, grammatical relations, and linear precedence on three different levels, in such a way that if the surface string is known, the parser can compute the semantics, and if the semantics is known, the same parser can compute the surface string(s).
We have tried this for examples of the complexity of "Peter likes Mary", but especially the semantic representation is still wide open. Furthermore, more complex examples run into limitations of XDG, which basically requires that the structures on each level have the same number of nodes. This might be solved by leaving nodes unconnected on some levels.
The end result we envision is a general way to do realization with XDG which has been checked by hand on some nontrivial examples, and which is ready to be implemented. There will probably not be enough time for the implementation itself.
We assume that the participants of this project have some familiarity with dependency grammars. It will be necessary to understand some details about modelling with XDG before the project starts. A good starting point is (Debusmann & Duchier, 2003) cited below. We will also have introductory sessions of some form in the first week of the EGK meeting, in order to bring everybody up to the necessary level of expertise.
Ralph Debusmann and Denys Duchier (2003). A Meta-Grammatical Framework for Dependency Grammar. Manuscript. http://www.ps.uni-sb.de/Papers/abstracts/DebusmannDuchierACL03.html
Ralph Debusmann rade@ps.uni-sb.de