nltk.inference.discourse

Module for incrementally developing simple discourses, and checking for semantic ambiguity, consistency and informativeness.

Many of the ideas are based on the CURT family of programs of Blackburn and Bos (see http://homepages.inf.ed.ac.uk/jbos/comsem/book1.html).

Consistency checking is carried out by using the mace module to call the Mace4 model builder. Informativeness checking is carried out with a call to Prover.prove() from the inference module.

DiscourseTester is a constructor for discourses. The basic data structure is a list of sentences, stored as self._sentences. Each sentence in the list is assigned a "sentence ID" (sid) of the form si. For example:

s0: A boxer walks
s1: Every boxer chases a girl

Each sentence can be ambiguous between a number of readings, each of which receives a "reading ID" (rid) of the form si -rj. For example:

s0 readings:

s0-r1: some x.(boxer(x) & walk(x))
s0-r0: some x.(boxerdog(x) & walk(x))

A "thread" is a list of readings, represented as a list of rids. Each thread receives a "thread ID" (tid) of the form di. For example:

d0: ['s0-r0', 's1-r0']

The set of all threads for a discourse is the Cartesian product of all the readings of the sequences of sentences. (This is not intended to scale beyond very short discourses!) The method readings(filter=True) will only show those threads which are consistent (taking into account any background assumptions).