Models for computer generated parody

Abstract

This paper outlines two approaches to the construction of computer systems that generate prose in the style of a given author. The first involves using intuitive notions of stylistic trademarks to construct a grammar that characterizes a particular author in this case, Ernest Hemingway. The second uses statistical methods for inferring a grammar from samples of an author's work in this instance, Thomas Hardy. A brief outline of grammar induction principles is included as background material for the latter system. The relative merits of each approach are discussed, and text generated from the resulting grammars is assessed in terms of its parodic quality. Further to its esoteric interest, a discussion of parody generation as a useful technique for measuring the success of grammatical inferencing systems is included, along with suggestions for its practical application in areas of language modeling and text compression.