yskip

C++ implementation of incremental skip-gram model with negative sampling (Kaji and Kobayashi, 2017).

Installation

% ./configure
% make
% make install

gcc 4.4.7 or higher is required.

Basic Usage

The following command learns skip-gram model from the plain text file text and save it as model:

% yskip text model

Learning from the standard input stream

yskip is able to learn from the standard input stream (without allocating the entire input on memory) by setting the input file name to -:

% cat text | yskip - model

This is usable only in online and mini-batch setting (see the options).

Options

yskip supports learning options listed below.

yskip [option] <train> <model>

Skip-gram model paramters:
 -d, --dimensionality-size=INT      Dimensionality of word embeddings (default: 100)
 -w, --window-size=INT              Window size (default: 5)
 -n, --negative-sample=INT          Number of negative samples (default: 5)
 -a, --alpha=FLOAT                  Distortion parameter (default: 0.75)
 -s, --subsampling-threshold=FLOAT  Subsampling threshold (default: 1.0e-3)
 -u, --unigram-table-size=INT       Unigram table size used for negative sampling (default: 1e8)
 -m, --max-vocabulary-size=INT      Maximum vocabulary size (default: 1e6)
 -e, --eta=FLOAT                    Initial learning rate of AdaGrad (default: 0.1)
 -b, --mini-batch-size=INT          Mini-batch size (default: 10000)
 -B, --binary-mode                  Read/write models in a binary format
 -l, --learning-strategy=INT        Learning strategy
                                    0: batch
                                    1: online
                                    2: mini-batch (default)
 -i, --iteration-numbedr            iteration number in batch learning (default: 5)

Misc.:
 -t, --thread-num=INT               Number of threads (default: 10)
 -I, --initial-model=FILE           Initial model (default: NULL)
 -r, --random-seed=INT              Random seed (default: current Unix time)
 -q, --quiet                        Do not show progress messages
 -h, --help                         Show this message

Learning strategy

yskip supports three learning strategies:

-l, --learning-strategy=INT        Learning strategy
                                    0: batch
                                    1: online
                                    2: mini-batch (default)

The batch strategy first calculates the noise distribution from the entire training data, and then performs mini-batch SGD to learn skip-gram model.
The online strategy performs the incremental SGNS described in (Kaji and Kobayashi, 2017).
The mini-batch updates the noise distribution every mini-batch, and performs mini-batch SGD to learn skip-gram model.

Incremental learning

yskip is able to incrementally update an old model, which is specified by using -I option, to obtain a new one:

% cat text-1 | yskip - model.1
% cat text-2 | yskip -I model.1 - model.2

This produces the same result as the following command (if random seed is set the same).

% cat text-1 text-2 | yskip - model.2

Converting the model file into word2vec-like format

The script perl/to_word2vec.pl converts yskip model file into word2vec-like format.

perl/to_word2vec.pl < model.yskip > model.w2v

Citation

Please cite (Kaji and Kobayashi, 2017) when using this code to publish a paper.

@inproceedings{kaji17,
  author    = {Nobuhiro Kaji and Hayato Kobayashi},
  title     = {Incremental Skip-gram Model with Negative Sampling},
  booktitle = {Proceedings of EMNLP},
  year      = {2017},
  pages     = {363--371}
}

arXiv version is also available here.

yahoojapan/yskip