A scalable framework needs to have the ways to actually increase coverage. The game heavily relies on prespecified resources including the grammar(s) and lexicon(s) of the NLP interface(s) as well as the knowledge base describing the game world and the related ontology. In its present state, these resources are manually compiled and are not extendable automatically by either online or offline techniques. The coverage of the game engine is then entirely dependent on their hard-wired substantial content.
A project concerning resources of the game might involve several things ranging from enhancing the robustness of the NLP interface as well as the reasoning modules, portability of the engine to other game domains. One important ingredient in this is the bootstrapping of resources. We think it might make an interesting project to see how far we can get in automating this bootstrapping procedure.
Bootstrapping could involve the creation of a world description of alternative game settings. This implies setting up a plotline as well supplying necessary domain ontology. Since this involves most of the hardest problems of AI, one can not hope to get much further than help reduce the efforts needed to dream up a new game. After automatic construction of at least a core, the toy-world description would need to be manually adjusted.
In the more interesting case, the game setting might be taken from a plot or nominal domain of a piece of natural language text, e.g., a children's fairy tale. In particular, characters, objects, places and the story's plotline could serve as the basis of the new game world. Shallow as well as deep NLP techniques have been proposed to attack such a knowledge acquisition problem. Shallow techniques (like text topic tiling and named entity recognition) use keyword distribution and shallow textual characteristics and could be utilized for character discovery or a rough description of spatial setting of the plot. Deep methods rely on linguistically parsed data to construct ontological knowledge. General world knowledge ontologies like WordNet, FrameNet, CYC or KT could then be used to help extend the knowledge base to suit the new domain.
The semi-automatic bootstrapping of the desription logic resources based on text has the additional advantage that extension of the natural language core lexicon resource 'comes for free'. However, a feasible restriction of bootstrapping could investigate the problem of lexicon scaling independently of the core game world. The extention of the lexicon would enable the engine to handle paraphrases of the same intended move. Ways of adding new lexical items could be explored to allow for alternative expressions of the main objects and eventualities of the game. This might also involve the scaling of knowledge base by including meaning pustulates and implicature to handle e.g., hyponymy. Existing lexical databases and the ontologies mentioned above are available to tackle such problems.
Victor Tron a1vtron@inf.ed.ac.uk