Applying information criteria in Skip-gram dimensionality selection

Implementation of Skip-gram Dimensionality Selection via information criteria (SNML, AIC, BIC).

Requirement

Please make sure your computer has installed these programs below:

python3.7
pip

Set up

• Install dependencies

pip install -r requirements.txt

Google cloud storage (GCS) set up

We train models on multiple servers but save the result on a GCS bucket. Please create an env.ini file to store access to the bucket. The env.ini file should be place in the root directory. The file should include content as following:

[GCS]
sync = no
project_id = xxx
bucket = xxx
app_credential = xxx

Configs:

• sync: set to no if you do not want to us GCS
• project_id, bucket: information of the bucket
• app_credential: path to json credential to access GCS

1. Preprocess data

1.1. Artificial data

Artificial data is generated using jupyter notebooks. Please refer to to notebooks below for the details of the data generation process.

• Data for original Skip-Gram model:

notebooks/Generate context distributions.jpynb

• Data for Skip-Gram Negative Sampling model:

notebooks/Generate context distributions - SGNS.jpynb

1.2. Text data

Run prepocess.py file to prepocess data. This file takes .txt file as input. Please remove special characters such as .,:? etc in the text file.
Parameters:

• input: input text file path
• output: output directory
• batch_size: batch size of the process
• window_size: the window size in Skip-Gram model Example:

python preprocess.py --input text8 --output data/text8 --batch_size 1000 --window_size 5

Others parameter for preprocessing such as subsampling threshold can be set in config.ini.

2. Train Skip-gram

Data after preprocess step can be use to train Skip-gram. Training commands are described as below:

2.1. Train original Skip-Gram model

Original Skip-Gram model should be trained using GPUs, we use tensorflow to train this model. Run tf_based/train.py to train this model.
Example:

python tf_based/train.py --input_path data/text8/ --batch_size 10 --output_path output/text8/ --epochs 1 --n_embedding 5

See config.ini and tf_based/train.py for more parameters settings.

2.2. Train Skip-Gram model with Negative Sampling

Skip-Gram Negative Sampling model is trained with numpy. Training process need context distribution to sample negative samples. Context distribution can be achieved by runing: utils/context_distribution_from_raw.py.
Run np_based/train.py to train this model.
Example:

python np_based/train.py --input_path data/text8/ --batch_size 10 --output_path output/text8/ --epochs 1 --n_embedding 5

See config.ini and np_based/train.py for more parameters settings.

3. Estimate AIC & BIC

Estimating AIC & BIC for original Skip-Gram and Skip-Gram Negative Sampling by following programs:

• original Skip-Gram:

python tf_based/run_aic_bic.py

• Skip-Gram Negative Sampling:

python np_based/run_aic_bic.py

See each python file for parameters setting.

4. Estimate SNML codelength

Estimating SNML for original Skip-Gram and Skip-Gram Negative Sampling by following programs:

• original Skip-Gram:

python tf_based/snml/tf_based/train_snml.py

• Skip-Gram Negative Sampling:

python np_based/train_snml.py

Parameters:

• model: output directory of trained model
• context_path: path to context distribution
• snml_train_file: data file, which each element will be estimated codelength
• scope: how many data elements will be estimated
• epochs: the number of epochs in gradient descent while training data in SNML
• n_context_sample: number of samples in importance sampling
• learning_rate: learning rate in gradient descent while training data in SNML
• continue_from: (do not set this in the first run) the number of element in the last run to continue
• continue_scope: (do not set this in the first run) continue from previous run, this parameter state number of scope in the previous run