nn-depparse

A Torch/Scala reimplementation of the neural network dependency parser described in Chen and Manning '14.

Requirements

• sbt 1.13.9
• torch
• torchx
• nn
• optim

Setup

1. Set up environment variables. For example, from the root directory of this project on blake:

export NNDEPPARSE_ROOT=`pwd`
export DATA_DIR=/iesl/canvas/strubell/data/

2. Put a word embeddings file in $NNDEPPARSE_ROOT/data/embeddings. The file is expected to contain one embedding per line, where the first field is the token and the remaining fields are the values of the embedding, each field separated by a single space. You can get the Collobert et al. embeddings here.
3. Compile: sbt compile
4. Perform all data preprocessing for a given configuration [also compiles]. For example:

./bin/all-data-processing.sh config/chen-ptb.conf

Running

1. Train the parser:

./bin/train-parser.sh config/chen-ptb.conf

2. Evaluate the parser (accuracy and speed):

./bin/parse-fast.sh config/chen-ptb.conf

3. Tune hyperparameters (assumes a GPU machine and uses all of its GPUs):

./bin/tune-hyperparams.sh config/chen-ptb.conf

[optional detail] Generating training data

1. Generate parse decisions + features for training from PTB: ./bin/get-parse-decisions-ptb.sh
2. Generate intmaps from parse decisions + features: ./bin/convert-ptb-feats-to-ints.sh
3. Generate Torch tensors from intmaps: ./bin/convert-ptb-feats-to-torch.sh
4. If word intmaps changed, generate Torch embedding tensors: ./bin/convert-collobert-embeddings-to-torch.sh

[optional detail] Generating test data

1. Generate dev/test intmaps for each sentence in PTB: ./bin/convert-ptb-sents-to-ints.sh
2. Generate Torch tensors from sentence intmaps: ./bin/convert-ptb-sents-to-torch.sh

[optional] Luajit hack to allow array-of-size def

In torch-distro/exe/luajit-rocks/luajit-2.1/src/luajit.c, add the function:

static int new_sized_table( lua_State *L )
{
    int asize = lua_tointeger( L, 1 );
    int hsize = lua_tointeger( L, 2 );
    lua_createtable( L, asize, hsize );
    return( 1 );
}

in main, after L is initialized add the lines:

lua_pushcfunction( L, new_sized_table );
lua_setglobal( L, "sized_table" );

Reinstall Torch.