A recurrent neural network that learns to define words from dictionaries. This repository is for AAAI2017 paper: "Definition Modeling: Learning to define word embeddings in natural language" along with the preporcessing scripts.

• Torch
• Python 2.7 (for basic scripts)
• Moses (specifically we just use sentence-bleu.cpp for evaluation)
• KenLM or SRILM (for ranking definition candidates, any ARPA file will do)
• Fish shell and Virtualenv (If you don't know them, you can skip. We have fish scripts help running python or torch batch commands)

Skip this if you do not have a GPU.

• CUDA Toolkit (v7.5)
• cuDNN (R5, only for convolution network)

Most of the libraries will come with Torch if you install from their installation script. You can use luarocks to install additional packages. For examples, luarocks install dp. The additional packages are:

• dp
• dpnn
• rnn

If you are planing to use GPU (CUDA), you will need the following packages:

• cutorch
• cunn
• cudnn (make sure that you get the right branch for your cuDNN version)

To install from source, go to the source code directory and run luarocks install.

• numpy
• KenLM (installation: pip install https://github.com/kpu/kenlm/archive/master.zip)

You will also need a set of word embeddings in torch binary format of an object:

{
  M, -- 2D tensor where each row is an embedding
  v2wvocab, -- index-to-word map
  w2vvocab -- word-to-index map
}

You can download embeddings from Word2Vec and use word2vec.torch to convert them into torch binary file.

In most of the scripts, there will be a help message which can be accessed by

th script.lua --help

• First you need to convert text data into torch binary files by using preprocess/prep_definition.lua. This will create multiple torch binary files in the data directory
• Then sub-select word embeddings using preprocess/prep_w2v.lua. This will align vocab and only save a set of embeddings we need)

We include our dataset (data/commondefs). If you want to use other dataset, please check the file format. For dictionary parsing scripts, check out dict-definition (only support WordNet and GCIDE for now).

• train.lua is a script for training a model
• test.lua is a script that uses a model to compute perplexity, generate definitions, and rank words (reverse dictionary).

Please see the option within the help message of the scripts.

• Add detail usage and examples to README
• Refactor ranking scripts