MUMBLE: A Flexible System for Language Production

MUMBLE Is a general purpose subsystem Intended as the final output segment of the natural language interface of an expert system. As described in [1], it takes plans expressed in nearly any conceptual representation and realizes them as English texts, integrating the different parts of the plan linguistically as well as taking full responsibility for the grammatically of the output and its coherency as a discourse. MUMBLE has been under development since 1975, having undergone four major overhauls, and is being used at MIT, BBN, and UMass. It is available from the author in a common dialect of maclisp and lispmachine lisp, with a translation to franz lisp underway and one to interllsp in planning. The kernal of MUMBLE consists of roughly 400 functions in eleven files, and as interpreted code occupies roughly 150k words of memory on a PDP/10. Runtimes are a function of the complexity of the input plans, but fall around .6 to 2 seconds per word of output text (again running Interpreted). Input plans for MUMBLE are invariably composed from the same data structures that the expert program uses Itself to represent what it wants to talk about. They are interpreted for production by a "dictionary" that one writes for each new expert program. This dictionary associates expressions from the expert with English phrases (or phrase schemas if the expression takes arguments); should more than one phrase be possible, one either includes a set of discriminating tests or makes use of the system supplied routine for managing synonyms. Example dictionaries are available for five different domains and their internal representations. MUMBLE supplies an extensive vocabulary of linguistic relations and operations in terms of which to write the needed English phrases, and at runtime maintains a thoroughly annotated abstract description of the linguistic structure already built from those elements of the plan that have been realized so far. This structure will automatically constrain the realization of the rest of the elements to insure their grammatlcality with respect to what has gone before; it also provides a context-sensitive reference for any heuristics one may write (or adapt from McDonald and Information Science Massachusetts those already included) to govern decisions such as the proper length of sentences, how to express contrasts, or how elaborate a description to use for some object. The paragraph below is an example of MUMBLE'S output, starting from the given list of assertions. (The task …