Talking to Ceptre: A Natural Language Interface

A natural language interface provides a simple way of communicating, by means of everyday language, with a computer. When you say "Hey Alexa", you are interfacing with the program code which controls Alexa's functionality by means of natural language. This is broadly what we wish to do in this thesis - to develop a method of interfacing with computers that is not dependent on one's knowledge of programming. Our interface supports the Ceptre programming language, developed by Chris Martens in the thesis "Programming Interactive Worlds with Linear Logic" [Martens, 2015] to allow one to formulate a narrative in terms of linear logic formulas by equating proof search with narrative description. Ceptre tackles problems surrounding the design of interactive worlds underlying narrative and games, and focuses on formalizing the notions of narrative generativity and causality. Such interactive worlds are shown to lend themselves to formal modeling by means of Jean-Yyves Girard's Linear Logic, and the correspondence may be extended to narratives in general. By approaching the problem of interactive game-design from the perspective of programming language design, Martens creates a new abstraction by means of which narrative structure can be understood. Our approach to the design of a natural language interface for Ceptre draws upon the correspondence architecture of LFG which establishes correspondences between different levels of linguistic representation. We will thus think of Ceptre program code as a level of linguistic representation similar to LFG c- and f-structures. This allows us to map from f-structures generated by LFG and a newly defined ?-structure, sets of which can be composed to form a Ceptre program. Underlying our approach is the claim by Richard Montague that "there is...no important theoretical difference between natural languages and the artificial languages of logicians" [Montague, 1970], which allows us to treat natural languages the same way as we do formal languages. Thus, though Ceptre is not a natural language, we can handle it in the same way as we do English. Our LFG f-structure gives us enough information to construct a semantic representation in the Ceptre programming language, which can then be used for inference as designed. The contribution of this thesis is thus this extension to the Ceptre language, a natural language interface which does not use statistical methods as is common in current day NLP research.