RelEx ( https://launchpad.net/relex ) produces semantic relations (expressed as word tuples) from natural language text. NLGEN does the inverse, generating natural language text from semantic relations in the same format.

The general idea is to parse English text using RelEx, produce new semantic relations using, for example, probabilistic inference, and then generate natural language based on them.

The basic strategy implemented in NLGEN is described in http://www.opencog.org/wiki/SegSim