Research at MIT
Robert C. Berwick · ACM SIGART Bulletin · 1982
A simple invertible grammar has been designed that is able to map input story summaries written in a simplified English (e.g., a plot outline of Macbeth, a medical history, or a synopsis of an historical event) into case-frame representations of those summaries. This input representation then serves as the basis for an analogical matching routine that is able to extract out the structural features common to several stories, a procedure that in turn forms the basis for an analogical learning program developed by Winston. The same grammar is then used in reverse by the analogical reasoner to generate natural sounding output. As implemented, the system is flexible and highly domainindependent; it relics on a modified system of order-independent transformations in order to obtain these properties. The generation procedure itself is broken down into three steps. Step I takes a highlevel plan generated by the analogy program and separates the plan into groups, generating a set of kernel sentences for each group. Step 2 takes as input the set of kernel sentences and a collection of syntactic and thematic rules and determines the set of transformations that must be applied upon each set of kernel sentences. Finally, Step 3 executes the transformations, combining the altered kernels of each set into a sentence, performs some clean-up pronominalization, and produces an output English word string. The following is an example of English text actually produced by the program: "At the beginning of the story Lady Macbeth is bothered by the fact that Macbeth is not the king. Macbeth who was a nobleman is persuaded by her to murder the king with a knife because she wants Macbeth to be the king."