Switch branches/tags
Find file History
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.
Tex
jars
scripts
src
trace
Makefile
README.md
active-arch.graffle
active-arch.png
en-cw.txt
init_train.conllx
init_train.conllx.bk
p3.pdf
test.conllx
test.conllx.bk
unlabeled_train.conllx
unlabeled_train.conllx.bk

README.md

How to run the program

You have several options to run the program. Please don't touch the shipped test.conllx.bk, unlabeled_train.conllx.bk, and init_train.conllx.bk. Those are test file, initial training files, and unlabeled training sentences pool. Those files can only be recreated via set.sh if you have access to UTCS lab machines.

Run all the experiments

  1. Type script to log all the terminal output (this will help to collect the full trace of training from stdout) and run make -j4. This probably the fastest way to run the experiments.

  2. Type script again and run make all. The advantage of this option is that all experiments are run sequentially and you may get a clean full trace in typescript file.

Run each individual experiment

  1. Type script and run:

    • random: make random
    • raw: make raw
    • length: make length
    • margin: make margin
  2. Run the Java program directly:

usage: java -cp stanford-corenlp-jar/*:src/*:. DependencyParserAPIUsage.java
 -embedFile <arg>       Path to embedding file
 -h,--help              Print out help manual
 -maxIter <arg>         maxIter property for Stanford Neural Network
                        Dependency Parser
 -model <arg>           Path where model is to be saved
 -numSentInInit <arg>   Number of sentences you want to use from initial
                        training set
 -numSentInPool <arg>   Number of sentences you want to pick from
                        "unlabeled" training pool in each iteration
 -outFile <arg>         Path where test data annotations are stored
 -output <arg>          Name of the file that you want to save the stdout
                        to
 -policy <arg>          Selection policy in the active learning
 -result <arg>          Name of the file that the result is to be saved
 -testFile <arg>        Path to test file
 -trainFile <arg>       Path to training file
 -unlabelTrain <arg>    Path to unlabeled training instances

For example, if I want to run the program with initial training set file path init_train.conllx, "unlabeled" training pool unlabeled_train.conllx, test set path test.conllx, number of sentences to use in the inital training set 2, and number of sentences to use in the "unlabeled" training set 1, I would run:

cp src/DependencyParserAPIUsage.java . ; \
javac -cp jars/*:src/*:. DependencyParserAPIUsage.java; \
java -cp jars/*:src/*:. DependencyParserAPIUsage \
                        -trainFile init_train.conllx \
                        -testFile test.conllx \
                        -embedFile en-cw.txt \
                        -model results/random_model \
                        -maxIter 500 \
                        -outFile annotations.conllx \
                        -unlabelTrain unlabeled_train.conllx \
                        -policy random \
                        -numSentInPool 1 \
                        -numSentInInit 2 \
                        -result results/result-500-random.txt &> results/result-random-500.log

Note

  1. Before running Makefile, make sure you want to backup any initial training file, "unlabeled" training pool, and test file with the file names with extra .bk extension in the same directory as your oginal files. make setup will assume the existence and the location of those backup files.

  2. Makefile current doesn't support -numSentInPool and -numSentInInit options. Run the Java program directly if you really want to use these two options.

  3. There are no safe guard against -numSentInPool and -numSentInInit. In other words, make sure you specify the number of sentences within the maximum sentences number of the files.

  4. To reproduce the experiments in the writeup, use the Makefile with default values.

  5. "results-batch2" is for the second run of experiments mentioned in the writeup. Results shown in Table 1 is from the ".txt" and ".log" files immediately under "trace" directory.

Directory Structure

.
├── emnlp2016.pdf
├── en-cw.txt
├── init_train.conllx
├── init_train.conllx.bk
├── jars
│   ├── commons-cli-1.4.jar
│   └── stanford-corenlp.jar
├── Makefile
├── README.md
├── scripts
│   ├── plot.py
│   ├── set.sh
│   └── split.awk
├── src
│   ├── DependencyParserAPIUsage.class
│   └── DependencyParserAPIUsage.java
├── test.conllx
├── test.conllx.bk
├── trace
│   ├── learning-curves
│   │   ├── curve.png
│   │   ├── curves-no-dash.png
│   │   ├── length.png
│   │   ├── margin.png
│   │   ├── random.png
│   │   └── raw.png
│   ├── result-500-length.log
│   ├── result-500-length.txt
│   ├── result-500-margin.log
│   ├── result-500-margin.txt
│   ├── result-500-random.log
│   ├── result-500-random.txt
│   ├── result-500-raw.log
│   ├── result-500-raw.txt
│   ├── results-batch2
│   │   ├── result-500-length.txt
│   │   ├── result-500-margin.txt
│   │   ├── result-500-random.txt
│   │   └── result-500-raw.txt
│   └── typescript
├── unlabeled_train.conllx
└── unlabeled_train.conllx.bk