Generalizing Syntactic Collocates for Creative Language Generation

This paper presents the construction of a data source that supports the automatic generation of cryptic crossword clues in a system called ENIGMA. Cryptic crossword clues have two layers of meaning: a surface reading that appears to be a fragment of English prose, and a puzzle reading that the solver must uncover to solve the clue. The content expressed by the clue, and the input to the generation process, is a word play puzzle, such as an anagram, perhaps. In expressing this puzzle ENIGMA must choose language creatively, so that a separate, surface reading of the text is also generated – in effect translating a semantic input via a layered text to a new semantic output. To ensure that this surface text is meaningful, ENIGMA uses corpus data to determine which words can be combined meaningfully and which cannot.