Code for generating synthetic PCFGs with desirable properties, that can be useful for learning experiments.

The generated PCFGs will be in CNF, and will be consistent (i.e. the sum of the probability of all strings will be 1.)

There are two main types that we generate: the first has a non trivial CFG backbone, and the second has a trivial CFG, 
including all possible productions.
The default settings will give grammars with 10 nonterminals, about 10,000 terminals, a length distribution that is 
zero truncated Poisson with expected length about 5 and a fat tailed lexical distribution (Zipfian) using a lognormal distribution.

python sample_grammar.py /tmp/grammar1.pcfg


python sample_fullgrammar.py /tmp/grammar1.pcfg


python sample_corpus.py /tmp/grammar1.pcfg /tmp/grammar1.samples


python sample_corpus.py --yieldonly --omitprobs /tmp/grammar1.pcfg /tmp/grammar1.samples