About

This is a repository containing the final project of the Deep Learning for Natural Language Processing course, part of the Cognitive Systems program and the Potsdam University.

The aim of the project is to reproduce the model and findings used in the "Enriching Word Vectors with Subword Information" by Bojanowski et al.

Authors of the project are Bhuvanesh Verma and Tamara Atanasoska.

Data

We used different version of Wikipedia data. Largest dataset we used is first 1 Billion bytes of English Wikipedia which can be obtained from here. Since this data dump is too big(around 1GB) it is not available in project repo. This data dump is raw web data which require some preprocessing(details can be found here).

For most of the training, we used smaller version of original data dump data/text8, which is one-seventh in size. Next dataset we used is Universal Dependencies dataset v2.8 for English(data/UD/UD_EN_train.txt) and German (data/UD/UD_DE_train.txt).

We used another set of datasets mentioned in Section 5.1 of "Enriching Word Vectors with Subword Information" to evaluate our model. These datasets are used for evaluating model on word similarity task.

English

1. eval/data/WS353/wordsim_relatedness_goldstandard.txt
2. eval/data/EN-RG-65.txt
3. eval/data/EN-RW.txt

German

1. eval/data/GUR65.txt
2. eval/data/GRU350.txt
3. eval/data/ZG222.txt

To evaluate model on word analogy task, we used the Google semantic/syntactic analogy datasets introduced in Mikolov et al. (2013) and its German translation.

English : eval/data/EN-GOOGLE.txt

German : eval/data/DE-GOOGLE.txt

Usage

In order to run the project, it is preferred to create a conda virtual environment. Start by installing requirements from requirements.txt as,

pip install -r requirements.txt

This will setup the environment to run project.

Run

Execute following command :

Train

python run.py --RUN_MODE train --SUBSAMPLING --NGRAMS --DATA data/text8

--RUN_MODE : {train, val}
--SUBSAMPLING : true if provided else false, use subsampling while training
--NGRAMS : true if provided else false, use ngrams while training
--DATA : path to data file
other arguments:
--VERSION : load a specific (saved) model using version number and checkpoint
--CKPT-E : checkpoint used to load a model to resume training or evaluate model
--RESUME : true if provided else false, resume from given epoch
--DEBUG  : to run overfitting

Resume

python run.py --RUN_MODE train --SUBSAMPLING --NGRAMS --DATA data/text8 --RESUME --VERSION VERSION_NUM --CKPT-E EPOCH_NUM

Evaluation

python run.py --RUN_MODE val --VERSION VERSION_NUM --CKPT-E EPOCH_NUM

Overfitting

python run.py --RUN_MODE train --DEBUG --SUBSAMPLING --NGRAMS --DATA data/text8

NOTE

For ngram model, we add embeddings of word and its ngrams. Resulting vector is then multiplied with target word embedding or negative word embeddings. This multiplication can result in large values and hence produce inf, -inf and nan values on performing further operations like taking sigmoid or log. To tackle this situation, we replace nan and inf values with 0 and max/min datatype values respectively using torch.nan_to_num() method. In addition to this, we normalize input and output embedding vectors which also helps in tackling this situation. Though after adding normalization, we required more training time to get better model.

Even after all these precautions, model can still produce nan value for loss during training. In this case, it is advised to decrease lr or batch_size.