Word Embedding


Gluon NLP makes it easy to evaluate and train word embeddings. Here are examples to evaluate the pre-trained embeddings included in the Gluon NLP toolkit as well as example scripts for training embeddings on custom datasets.

Word Embedding Evaluation

To evaluate a specific embedding on one or multiple datasets you can use the included evaluate_pretrained.py as follows.

$ python evaluate_pretrained.py

Call the script with the –help option to get an overview of the supported options. We include a run_all.sh script to run the evaluation for the pre-trained English Glove and fastText embeddings included in GluonNLP.

$ run_all.sh

The resulting logs and a notebook containing a ranking for the different evaluation tasks are available here.

Word Embedding Training (Skipgram and CBOW)

Besides loading pre-trained embeddings, the Gluon NLP toolkit also makes it easy to train embeddings.

The following code block shows how to use Gluon NLP to train a SkipGram or CBOW models. The script and parts of the Gluon NLP library support just-in-time compilation with numba, which is enabled automatically when numba is installed on the system. Please pip install –upgrade numba to make sure training speed is not needlessly throttled by Python.

$ python train_sg_cbow.py --model skipgram --ngram-buckts 0  # Word2Vec Skipgram
$ python train_sg_cbow.py --model skipgram --ngram-buckts 2000000  # fastText Skipgram
$ python train_sg_cbow.py --model cbow --ngram-buckts 0  # Word2Vec CBOW
$ python train_sg_cbow.py --model cbow --ngram-buckts 2000000  # fastText CBOW

Word2Vec models were introduced by Mikolov et al., “Efficient estimation of word representations in vector space” ICLR Workshop 2013. FastText models were introudced by Bojanowski et al., “Enriching word vectors with subword information” TACL 2017.

We report the results obtained by running the python3 train_sg_cbow.py --batch-size 4096 --epochs 5 --data fil9 --model skipgram script.For comparison we also report the results obtained by training FastText with the facebookresearch/fastText implementation. All results are obtained by training 5 epochs on the Fil9 dataset.

Similarity Dataset facebookresearch/fastText train_sg_cbow.py
WordSim353-similarity 0.752 0.734
WordSim353-relatedness 0.612 0.608
MEN (test set) 0.736 0.700
RadinskyMTurk 0.687 0.655
RareWords 0.420 0.457
SimLex999 0.320 0.346
SimVerb3500 0.190 0.235
SemEval17Task2 (test set) 0.541 0.542
BakerVerb143 0.406 0.383
YangPowersVerb130 0.489 0.466
Google Analogy Dataset facebookresearch/fastText train_sg_cbow.py
capital-common-countries 0.796 0.581
capital-world 0.442 0.334
currency 0.068 0.074
city-in-state 0.198 0.076
family 0.498 0.593
gram1-adjective-to-adverb 0.377 0.688
gram2-opposite 0.343 0.693
gram3-comparative 0.646 0.868
gram4-superlative 0.510 0.757
gram5-present-participle 0.445 0.792
gram6-nationality-adjective 0.828 0.840
gram7-past-tense 0.385 0.380
gram8-plural 0.706 0.810
gram9-plural-verbs 0.501 0.813

Loading of fastText models with subword information

Fasttext models trained with the library of facebookresearch are exported both in a text and a binary format. Unlike the text format, the binary format preserves information about subword units and consequently supports computation of word vectors for words unknown during training (and not included in the text format). Besides training new fastText embeddings with Gluon NLP it is also possible to load the binary format into a Block provided by the Gluon NLP toolkit using FasttextEmbeddingModel.load_fasttext_format.

Word Embedding Training (GloVe)

Gluon NLP also supports training GloVe models.

$ python train_glove.py tools/build/cooccurrences.npz tools/build/vocab.txt

Where the cooccurrences.npz is a numpy archive containing the sparse word-word cooccurrence matrix and vocab.txt a textfile containing the words and their counts. They can be constructed from a text corpus using the included vocab_count and cooccur tools. They can be used as follows

$ mkdir tools/build; cd tools/build; cmake ..; make
$ ./vocab_count corpus-part1.txt corpus-part2.txt > vocab.txt
$ ./cooccur corpus-part1.txt corpus-part2.txt < vocab.txt

Also see ./vocab_count –help and ./cooccur –help for configuration options such as min-count or window-size.

GloVe models were introduced by Pennington et al., “Glove: global vectors for word representation”, ACL 2014.