Graph-based Posterior Regularization for Structured Prediction
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
data
libs
scripts
src
supplementary
README.md
README.txt
build.xml

README.md


pr-graph version 0.1 (Graph-based Posterior Regularization)

This Javaproject implements the Graph-based Posterior Regularization model described in the following paper:

Graph-Based Posterior Regularization for Semi-Supervised Structured Prediction Luheng He, Jennifer Gillenwater, and Ben Taskar. Conference on Computational Natural Language Learning (CoNLL), 2013.


Contents

  1. Compiling
  2. Graph Building a. POS Tagging Graph b. Handwriting Letters Graph
  3. Running a. Input data format b. Running PR-graph

  1. Compiling

The build.xml is included in the repository. Use Ant (http://ant.apache.org/) to compile the project.

From the base directory of this project, run "ant build" to compile the code. You can also run "ant clean" to remove compiled files and start over.


  1. Graph Buiding


2a. POS Tagging Graph

PosTagging Graph building uses a suffix dictionary included in //pr-graph/data/suffix.dict This list of suffixes is extracted from the Wiktionary data.

Universal part-of-speech tags mapping can be found here: https://code.google.com/p/universal-pos-tags/

(D. Das, S. Petrov, and R. McDonald. 2012. A Universal Part-of-Speech Tagset. In Proc. LREC.)

To run the graph builder, we can do:

export WDIR="your working directory" export DDIR=”your data directory” export CLASSPATH="$WDIR/bin/:$WDIR/libs/optimization-2010.11.jar:$WDIR/libs/trove-2.0.2.jar:$WDIR/libs/args4j-2.0.10.jar"

java -cp $CLASSPATH -Xmx8000m programs.TestPosGraphBuilder
-data-path "$DDIR/lang.train,$DDIR/lang.test" \ # a list of comma-delimited input file paths -sufix-path “$DDIR/suffix.dict” -umap-path "$DDIR/lang.map"
-graph-path "$DDIR/graph/lang.grph" \ -ngram-path "$DDIR/graph/$lang.idx" \ -num-neighbors 60
-lang-name "lang"

The Graph builder outputs the node index file to -ngram-path, and the graph edge file to -graph-path. More options can be found at config.Config, config.PosConfig and config.PosGraphConfig.


2b. Handwriting Letters Graph

The code for building OCR Graph lives in another project (due to its dependency on the FastEMD code). The code is under //pr-graph/supplementary. We can also use the graph file in //pr-graph/data/graph to run the experiments.

The FastEMD code and its Java wrapper is written by Ofir Pele: (O. Pele and M. Werman. 2009. Fast and Robust Earth Mover’s Distances. In Proc. ICCV)

To run the OCR graph builder, we can do:

1). Extract the package ocr-graph-builder.tar.gz 2). In the file ocr-graph/src/ emd_hat.java: In Line 119, update the path of the emd tool library: System.load("/$YOUR_WORKING_PATH/ocr-graph/libs/libemd_hat_native.so"); 3). From the base directory ocr-graph, run: ant build to compile. 4). Run:

export WDIR="your working directory" export DDIR="your data directory" export CLASSPATH="$WDIR/bin/:$WDIR/libs/trove-2.0.2.jar:$WDIR/libs/args4j-2.0.10.jar:$WDIR/libs/libemd_hat_native.so"

java -cp $CLASSPATH -Xmx8000m OcrGraphBuilder -data-path "$DDIR/letter.data"


  1. Running


3a. Input data format

We use the CoNLL-X (http://ilk.uvt.nl/conll/index.html#dataformat) format for POSTagging, and the OCR (http://www.seas.upenn.edu/~taskar/ocr/) data for the handwriting task.


3b. Running PR-graph

For Pos-Tagging, run: java -cp $CLASSPATH -Xmx8000m programs.TestHighOrderPos -num-labels 100
-data-path "lang.train,lang.test"
-umap-path "lang.map"
-ngram-path "lang-graph.idx"
-graph-path "lang-graph.grph"
-lang-name "lang" -sample-fold 0
-num-sample-folds 10
-eta 0.2
-backoff 1e-8
-gaussian-prior 100
-graph-strength 0.1
-num-mstep-iters 300
-num-estep-iters 10
-em-stop 0.01
-estep-stop 0.01
-mstep-stop 0.00001
-num-em-iters 20
-num-threads 8
-encoding "LATIN1"

*About encoding: There was a encoding bug in the code for CoNLL-2013 paper, so in order to reproduce the result exactly, set -encoding to "LATIN1"; otherwise, set the encoding to "UTF8". The difference is tiny.

*Numerical issue in multi-threading There will be tiny difference in optimizing for the CRF base model (probably due to some numerical problem) when we change the number of threads. Use -num-thread=8 to reproduce result. I will try to fix this problem in future version.

*-sample-fold and -num-sample-folds In the CoNLL 2013 paper, we ran the experiment 10 times by randomly sampling 10 different set of training samples, so we used -num-sample-folds = 10. -sample-fold=0 means it is using the 0-th training set. Averaging through sample-fold from 0 to 9 will get the final results.


For handwriting recognition, run:

java -cp $CLASSPATH -Xmx8000m programs.TestHighOrderOCR -num-labels 110
-data-path "letter.data"
-graph-path "ocr-graph.grph"
-sample-fold 0
-num-sample-folds 10
-eta 0.2
-backoff 1e-8
-gaussian-prior 100
-graph-strength 1.0
-num-mstep-iters 300
-num-estep-iters 10
-em-stop 0.01
-estep-stop 0.01
-mstep-stop 0.00001
-num-em-iters 20
-num-threads 8 \