PERPL stands for Probabilistic Exact Recursive Programming Language. Unlike many general-purpose PPLs, PERPL performs exact inference, via factor graph grammars. And unlike other PPLs that do support exact inference, PERPL can express unbounded recursive calls and (with some restrictions) recursive data structures. For example, a PCFG parser written in PERPL appears to generate (infinitely many) trees and sum the probabilities of those (exponentially many) trees that yield a given string; yet it compiles to a cubic-sized system of equations whose solution is equivalent to the CKY algorithm.
For more about the language, see language.md.
To build the compiler (requires GHC):
make
To run tests:
make tests
To compile a PERPL program:
./perplc [options] FILE.ppl
Options:
-O0 -O1 Optimization level (0 = off, 1 = on)
-l Don't linearize the file (implies -c -e)
-e Don't eliminate recursive datatypes (implies -c)
-d DTYPES Defunctionalize recursive datatypes DTYPES
-r DTYPES Refunctionalize recursive datatypes DTYPES
-c Compile only to PPL code (not to FGG)
-z Compute sum-product
-o OUTFILE Output an FGG to OUTFILE
Although perplc can compute sum-products, its implementation is not very efficient. The normal usage is to use the -o option to output an FGG to a JSON file. Then the FGG can be processed using the fggs package.
This code is written by Colin McDonald (University of Notre Dame) with contributions from David Chiang (University of Notre Dame) and Chung-chieh Shan (Indiana University). It is licensed under the MIT License.