GA-FLM is a genetic-based algorithm for optimizing the structure of Factored Language Models (FLMs). It is used as an extension to the FLM programs in the SRI Language Modeling toolkit.

The program takes as input some training/development text files and some parameter files that specify the type of genetic algorithms and factored language model desired by the user. It then uses standard genetic algorithms search to build a population of factored language models. The search optimizes for development set perplexity. GA-FLM is most useful when one has many factors specified for language model development, but cannot find a good factored language model manually.

Terms of Use

GA-FLM can be downloaded free of charge under the GNU Public Licence. The current download is a beta version (v.0.1). For bugs, suggestions, or questions, please feel free to contact Kevin Duh (duh at ee dot washington.edu). The code is still under development, so I am more than happy to receive suggestions and questions.

Published research using GA-FLM may cite the following paper:

• Kevin Duh and Katrin Kirchhoff, "Automatic Learning of Language Model Structure", Proc. of the 20th International Conference on Computational Linguistics (COLING-2004), Geneva, Switzerland, August 2004.

Download

• Download the tar-zipped source code.
• The README and user manual, which are also included in the above package.
• A copy of the COLING paper and UWEE technical report associated with GA-FLM.
• The SRI Language Modeling toolkit is required for GA-FLM. Be sure to use SRILM version 1.4.1 or above, as it contains the correct FLM programs.