This is an implementation of the LexVec word embedding model (similar to word2vec and GloVe) that achieves state of the art results in multiple NLP tasks, as described in these papers.

• Common Crawl - 58B tokens, cased - 2,000,000 words - 300 dimensions
  • Word Vectors (2.1GB)
  • Binary model (8.6GB) - use this to compute vectors for out-of-vocabulary (OOV) words

• Common Crawl - 58B tokens, lowercased - 2,000,000 words - 300 dimensions

  • Word Vectors (2.2GB)
  • Word + Context Vectors (2.3GB)

• English Wikipedia 2015 + NewsCrawl - 7B tokens - 368,999 words - 300 dimensions

  • Word Vectors (398MB)
  • Word + Context Vectors (426MB)

• Both models use vectors with 300 dimensions.

• Both models use character n-grams of length 3-6 as subwords.

• All tasks are evaluated using cased words ("Toronto" != "toronto").

• GSem, GSyn, and MSR analogies were solved using 3CosMul.

• Both models were trained using this release of Common Crawl which contains 58B tokens, restricting the vocabulary to the 2 million most frequent cased words.

• Subword LexVec was trained using the following command:

  $ OUTPUT=output scripts/em_lexvec.sh -corpus common_crawl_cased.txt -negative 3 -dim 300 -subsample 1e-5 -minfreq 0 -window 2 -minn 3 -maxn 6
  

• fastText was trained using the following command:

  $ ./fasttext skipgram -input common_crawl_cased.txt -minCount 0 -t 1e-5 -dim 300 -lr 0.025 -minn 3 -maxn 6                  
  

• All three models were trained using the same English Wikipedia 2015 + NewsCrawl corpus.

• GSem, GSyn, and MSR analogies were solved using 3CosMul.

• LexVec was trained using the default parameters, expanded here for comparison:

  $ OUTPUT=output scripts/im_lexvec.sh -corpus enwiki+newscrawl.txt -dim 300 -window 2 -subsample 1e-5 -negative 5 -iterations 5 -minfreq 100 -model 0 -minn 0
  

• word2vec Skip-gram was trained using:

  $ ./word2vec -train enwiki+newscrawl.txt -output sgnsvectors -size 300 -window 10 \
  -sample 1e-5 -negative 5 -hs 0 -binary 0 -cbow 0 -iter 5 -min-count 100
  

• All models use vectors with 300 dimensions.

• GSem, GSyn, and MSR analogies were solved using 3CosMul.

• LexVec was trained using this release of Common Crawl which contains 58B tokens, restricting the vocabulary to the 2 million most frequent words, using the following command:

  $ OUTPUT=output scripts/em_lexvec.sh -corpus common_crawl.txt -negative 3 -dim 300 -subsample 1e-5 -minfreq 0 -window 2 -minn 0                                     
  

• The pre-trained word2vec vectors were trained using the unreleased Google News corpus containing 100B tokens, restricting the vocabulary to the 3 million most frequent words.

• The pre-trained GloVe vectors were trained using Common Crawl (release unknown) containing 42B tokens, restricting the vocabulary to the 1.9 million most frequent words.

1. Install the Go compiler and clang.

2. Make sure your $GOPATH is set

3. Execute the following commands in your terminal:

   $ go get github.com/alexandres/lexvec
   $ cd $GOPATH/src/github.com/alexandres/lexvec
   $ make

To get started, run $ scripts/demo.sh which trains a model using the small text8 corpus (100MB from Wikipedia).

Basic usage of LexVec is:

$ OUTPUT=dirwheretostorevectors scripts/im_lexvec.sh -corpus somecorpus

Run $ ./lexvec -h for a full list of options.

By default, LexVec stores the sparse matrix being factorized in-memory. This can be a problem if your training corpus is large and your system memory limited. We suggest you first try using the in-memory implementation. If you run into Out-Of-Memory issues, use the External Memory variant with the -memory option specifying how many GBs of memory to use for the sort buffer.

$ OUTPUT=dirwheretostorevectors scripts/em_lexvec.sh -corpus somecorpus -memory 4. ...exactsameoptionsasinmemory

Subword information is controlled by the options -minn, -maxn, and -subword.

• To disable the use of subword information, specify -minn 0.

• To use character n-grams of length 3-6, specify -minn 3 -maxn 6 (this is the default configuration).

• To provide your own subword information (such as morphological segmentation), specify -minn 0 -subword subwords.txt, where the subwords file contains one line for each vocabulary word (vocabulary must match that of -vocab), each line containing a word followed by each of its subwords, separated by spaces.

By default, the binary model used for computing OOV word vectors is saved to $OUTPUT/model.bin. Set -outputsub "" to disable saving this model.

Use the binary model to compute vector for OOV words:

• Using the Go executable by providing one word per line on stdin:

  $ echo "marvelicious" | ./lexvec embed -outputsub pathtomodel.bin

• Using the Python lib:

  import lexvec
  model = lexvec.Model('pathtomodel.bin')
  vector = model.word_rep('marvelicious')

Note: You can also use these commands to get vectors for in-vocabulary words as the binary model stores the vocabulary used for training.

Alexandre Salle, Marco Idiart, and Aline Villavicencio. "Matrix Factorization using Window Sampling and Negative Sampling for Improved Word Representations." ACL (2016). (pdf)

Alexandre Salle, Marco Idiart, and Aline Villavicencio. "Enhancing the LexVec Distributed Word Representation Model Using Positional Contexts and External Memory." arXiv preprint arXiv:1606.01283 (2016). (pdf)

Alexandre Salle and Aline Villavicencio. "Incorporating Subword Information into Matrix Factorization Word Embeddings." Second Workshop on Subword and Character LEvel Models in NLP (2018). (pdf)

Alexandre Salle and Aline Villavicencio. "Why So Down? The Role of Negative (and Positive) Pointwise Mutual Information in Distributional Semantics." arXiv preprint arXiv:1908.06941 (2019). (pdf)

Copyright (c) 2016-2018 Salle, Alexandre alex@alexsalle.com. All work in this package is distributed under the MIT License.