Skip to content


Repository files navigation

#Generalized Language Model Toolkit

The software can be used to compute a Generalized Language Model which is yet another mean to compute a Language Model. As shown in this publication Generalized Language models can outperform Modified Kneser Ney Smoothing by 10 to 25 % in Terms of perplexity.

Getting started

git clone
sudo chmod a+x

You will need to install maven in order to build the project.

sudo apt-get install maven2

You need to copy config.sample.txt to config.txt and read the instructions in config.sample.txt.

cp config.sample.txt config.txt
emacs config.txt

After you set all your directories in config.txt you can run the project


Disk and Main memory requirements

Since Generalized language models can become very large the software is written to use the hard disk. In this sense you can theoretically run the programm with very little memory. Still we recommend 16 GB of main memory for the large english wikipedia data sets.

We tried to avoid frequent disc hits. Still the programm will execute much faster if you store your data on a Solid State disk.

Download the test data sets

you need to have a file called normalized.txt which serves as your input. This file should contain one sentence per line. You will learn language models based on this file.

Please refere to in order to download preprocessed and formatted data sets.

If you whish to parse the data yourself (e.g. because you want to use a newer wikipedia dump) refer to

Processing pipeline of the GLM toolkit:

you have to start with a file called normalized.txt which has to be stored in your data directory (according to config.txt). will compile the program and start the flow of the following steps (which can be configured by switching the fields ind config.txt from true to false)

  • splitting normalized.txt to training.txt and testing.txt according to the datasplit parameters in config.txt
  • building a wordindex index.txt this index is used to split the language models into files of equal size
  • creating absolute counts and continuation counts in the directories absolute and continuation ** the various models are stored in folders like 11111 meaning a regular 5 gram or 11011 meaning a skipped 5 gram at the third position
  • creation of testing samples from testing.txt: testing-samples-4.txt for example contains about 100k sequences of 4 words to be tested
  • calculating the D and N values for Modified Kneser Ney Smoothing and making them persistent in the two *.ser files (for speeding up various tests)
  • running the experiments by creating files like mod-kneser-ney-complex-backoffToCont-3.txt: depending on your configuration the files could be named with a simple instead of complex (complex meaning GLM, simple meaning LM). Exchanging the 3 you can have different model lenghts. These files contain the testing samples with the log of their probabilities.
  • you have to manually calculate the entropy by running the python script as an argument you might want to pass mod*.txt: in this way you can calculate the entropy for all files and experiments.

Citing the paper

If this software or data is of any help to your research please be so fair and cite the original publication which is also in the home directory of [this git repository]( Generalized Language Model as the Combination of Skipped n-grams and Modified Kneser-Ney Smoothing.pdf). You might want to use the following bibtex entry:

   author = {Pickhardt, Rene and Gottron, Thomas and Körner, Martin and  Wagner, Paul Georg and  Speicher, Till and  Staab, Steffen}, 
   title = {A Generalized Language Model as the Combination of Skipped n-grams and Modified Kneser Ney Smoothing}, 
   year = {2014}, 
   booktitle = {ACL'14: Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics}, 


The Generalized Language models envolved from Paul Georg Wagner's and Till Speicher's Young Scientists project called Typology which I advised in 2012. The Typology project played around and evaluated an idea I had (inspired by the PhD thesis of Adam Schenker) of presenting text as a graph in which the edges would encode relationships (nowerdays known as skipped bi-grams). The Graph was used to produce an answer to the next word prediction problem applied to word suggestions in keyboards of modern smartphones. From the convincing results I developed the theory of Generalized Language models. Most of the Code was written by my student assistent Martin Körner who also created his bachlor thesis about the implementation of a preliminary vesion of the Generalized Language Models. This thesis is a nice reference if you want to get an understanding of modified kneser ney smoothing for standard language models. In terms of notation and building of generalized language models it is outdated.

Questions, Feedback, Bugs

If you have questions feel free to contact me via the issue tracker. on my blog or in the paper you could find my mail address.