An implementation of parallel gibbs sampling for word segmentation and POS tagging.
Branch: master
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Type Name Latest commit message Commit time
Failed to load latest commit information.

pgibbs - Parallelized Unsupervised word segmentation and POS tagging

by Graham Neubig

This package implements parallel Gibbs sampling for word segmentation and POS tagging using both blocked and multi-sampler methods. The two executables are src/bin/pgibbs-ws for word segmentation and src/bin/pgibbs-hmm for POS tagging. You can find more details in the following paper:

"Simple, Correct Parallelization for Blocked Gibbs Sampling". Graham Neubig. Technical Report. 2014.

  author = {Graham Neubig},
  title = {Simple, Correct Parallelization for Blocked Gibbs Sampling},
  institution = {Nara Institute of Science and Technology},
  year = {2014},
  url = {}


To compile pgibbs, download it from git, and run

autoreconf -i

Data Preparation

For HMM-based tagging, the input is divided into words:

上海 浦东 开发 与 法制 建设 同步

And for word-segmentation, the input is divided into characters:

上 海 浦 东 开 发 与 法 制 建 设 同 步

If you want to replicate experiments in the technical report, you must have the Chinese Treebank 5.0, which can be obtained from the LDC (details here).

Running Programs

Both programs can be run with


The input data is the previously mentioned files, and the output data is a prefix where the label files will be written. The major options used are as follows:

-iters 2000             // The number of iterations
-threads {1,2,4,8}      // The number of threads to use
-blocksize {1,2,4,10,20,40,100} // The size of a single block for blocked sampling
-sampmeth {block,parallel} // Whether to perform blocked or parallel sampling
-skipiters {0,2000}     // Will skip a Metropolis-Hastings rejection for a certain
                        //  number of iterations, when set equal to the number of
                        //  iterations, MH will not be performed

One thing to note here is that if you are using "-sampmeth block", "-blocksize" should be larger than "-threads", maybe 4 times larger (see the paper for details).

There are also additional options for word segmentation only:

-n 2                    // The n-gram size of the model

Or for the HMM only:

-classes 30             // The number of classes in the model

Execution Examples

An example of running the program for part-of-speech tagging for the file test.words on 4 cores is below:

pgibbs-hmm -iters 500 -threads 4 -blocksize 16 -sampmeth block -skipiters 500 test.words output-hmm

An example of running the program for word segmentation on the file test.chars on 4 cores is below:

pgibbs-ws  -iters 500 -threads 4 -blocksize 16 -sampmeth block -skipiters 0   test.chars output-ws