This is an archive of the code used for the paper Learning Distributed Representations for Structured Output Prediction that appeared at NIPS 2014.

Getting started

To compile, you need scala and SBT installed on your computer. The newsgroup experiments were run with at least 15 GB of RAM. The POS experiments required much more RAM (over 40GB).

First, compile by running

sbt compile

If you want to clean up all the compiled files, delete the target, lib_managed and project directories.

Replicating newsgroup results

To train a multiclass classifier using data that is formatted in the lib-linear data format, use ./run.sh linear. Running this should list the different command line switches that you can use. The most important switches are:

1. -n: The dimensionality of the label vectors. If this dimensionality is more than the number of labels in the problem, then the dimensionality is set to the number of labels.
2. -w: Use one-hot vectors only (i.e. do not train label vectors)

See the documentation for an explanation of the other options. For every run, the complete log of the execution will be saved in a subdirectory of the directory experiments.

To produce the newsgroup results, the following settings were used. First, the data:

TRAIN_FILE=data/20news/features.extracted/20news-bydate-train-stanford-classifier.txt.feats
TEST_FILE=data/20news/features.extracted/20news-bydate-test-stanford-classifier.txt.feats

1. Baseline: Structured SVM

   ./run.sh linear -n 21 \
            -w \
            --train-iters 25 \
            -t $TRAIN_FILE \
            -e $EVAL_FILE \
            -v --cv-iters 5
       

2. DISTRO

   RANDOM_SEED=1 && ./run.sh linear -n 20 \
                   -a l2-prox-alternating \
                   --train-iters 20 \
                   -t $TRAIN_FILE \
                   -e $TEST_FILE \
                   --weight-train-iters 5 \
                   --label-train-iters 5 \
                   --lambda1 0.25 \
                   --lambda2 0.004096
       

   To find the best parameters by cross validation on the training set, you can remove the --lambda1 and --lambda2 options and replace them with a -v option to indicate five fold cross validation (as in the baseline). The parameters in the command above are the result of cross-validation.

Replicating POS results

To replicate the POS tagging results, you need to access the correct sections of the Penn Treebank and the Basque data in CONLL format. Use ./run.sh pos and ./run.sh conll.pos respectively to get information about the command line options. (The options are very similar to the newsgroup ones.)