gensim logo

gensim tagline

Get Expert Help

• machine learning, NLP, data mining

• custom SW design, development, optimizations

• corporate trainings & IT consulting

models.word2vec – Deep learning with word2vec

models.word2vec – Deep learning with word2vec

Deep learning via word2vec’s “skip-gram and CBOW models”, using either hierarchical softmax or negative sampling [1] [2].

The training algorithms were originally ported from the C package and extended with additional functionality.

For a blog tutorial on gensim word2vec, with an interactive web app trained on GoogleNews, visit

Make sure you have a C compiler before installing gensim, to use optimized (compiled) word2vec training (70x speedup compared to plain NumPy implementation [3]).

Initialize a model with e.g.:

>>> model = Word2Vec(sentences, size=100, window=5, min_count=5, workers=4)

Persist a model to disk with:

>>> model = Word2Vec.load(fname)  # you can continue training with the loaded model!

The model can also be instantiated from an existing file on disk in the word2vec C format:

>>> model = Word2Vec.load_word2vec_format('/tmp/vectors.txt', binary=False)  # C text format
>>> model = Word2Vec.load_word2vec_format('/tmp/vectors.bin', binary=True)  # C binary format

You can perform various syntactic/semantic NLP word tasks with the model. Some of them are already built-in:

>>> model.most_similar(positive=['woman', 'king'], negative=['man'])
[('queen', 0.50882536), ...]

>>> model.doesnt_match("breakfast cereal dinner lunch".split())

>>> model.similarity('woman', 'man')

>>> model['computer']  # raw numpy vector of a word
array([-0.00449447, -0.00310097,  0.02421786, ...], dtype=float32)

and so on.

If you’re finished training a model (=no more updates, only querying), you can do

>>> model.init_sims(replace=True)

to trim unneeded model memory = use (much) less RAM.

Note that there is a gensim.models.phrases module which lets you automatically detect phrases longer than one word. Using phrases, you can learn a word2vec model where “words” are actually multiword expressions, such as new_york_times or financial_crisis:

>>> bigram_transformer = gensim.models.Phrases(sentences)
>>> model = Word2Vec(bigram_transformer[sentences], size=100, ...)
[1]Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. Efficient Estimation of Word Representations in Vector Space. In Proceedings of Workshop at ICLR, 2013.
[2]Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg Corrado, and Jeffrey Dean. Distributed Representations of Words and Phrases and their Compositionality. In Proceedings of NIPS, 2013.
[3]Optimizing word2vec in gensim,
class gensim.models.word2vec.BrownCorpus(dirname)

Bases: object

Iterate over sentences from the Brown corpus (part of NLTK data).

class gensim.models.word2vec.LineSentence(source, max_sentence_length=10000, limit=None)

Bases: object

Simple format: one sentence = one line; words already preprocessed and separated by whitespace.

source can be either a string or a file object. Clip the file to the first limit lines (or no clipped if limit is None, the default).


sentences = LineSentence('myfile.txt')

Or for compressed files:

sentences = LineSentence('compressed_text.txt.bz2')
sentences = LineSentence('compressed_text.txt.gz')
class gensim.models.word2vec.Text8Corpus(fname, max_sentence_length=10000)

Bases: object

Iterate over sentences from the “text8” corpus, unzipped from .

class gensim.models.word2vec.Vocab(**kwargs)

Bases: object

A single vocabulary item, used internally for collecting per-word frequency/sampling info, and for constructing binary trees (incl. both word leaves and inner nodes).

class gensim.models.word2vec.Word2Vec(sentences=None, size=100, alpha=0.025, window=5, min_count=5, max_vocab_size=None, sample=0.001, seed=1, workers=3, min_alpha=0.0001, sg=0, hs=0, negative=5, cbow_mean=1, hashfxn=<built-in function hash>, iter=5, null_word=0, trim_rule=None, sorted_vocab=1, batch_words=10000)

Bases: gensim.utils.SaveLoad

Class for training, using and evaluating neural networks described in

The model can be stored/loaded via its save() and load() methods, or stored/loaded in a format compatible with the original word2vec implementation via save_word2vec_format() and load_word2vec_format().

Initialize the model from an iterable of sentences. Each sentence is a list of words (unicode strings) that will be used for training.

The sentences iterable can be simply a list, but for larger corpora, consider an iterable that streams the sentences directly from disk/network. See BrownCorpus, Text8Corpus or LineSentence in this module for such examples.

If you don’t supply sentences, the model is left uninitialized – use if you plan to initialize it in some other way.

sg defines the training algorithm. By default (sg=0), CBOW is used. Otherwise (sg=1), skip-gram is employed.

size is the dimensionality of the feature vectors.

window is the maximum distance between the current and predicted word within a sentence.

alpha is the initial learning rate (will linearly drop to min_alpha as training progresses).

seed = for the random number generator. Initial vectors for each word are seeded with a hash of the concatenation of word + str(seed). Note that for a fully deterministically-reproducible run, you must also limit the model to a single worker thread, to eliminate ordering jitter from OS thread scheduling. (In Python 3, reproducibility between interpreter launches also requires use of the PYTHONHASHSEED environment variable to control hash randomization.)

min_count = ignore all words with total frequency lower than this.

max_vocab_size = limit RAM during vocabulary building; if there are more unique words than this, then prune the infrequent ones. Every 10 million word types need about 1GB of RAM. Set to None for no limit (default).

sample = threshold for configuring which higher-frequency words are randomly downsampled;
default is 1e-3, useful range is (0, 1e-5).

workers = use this many worker threads to train the model (=faster training with multicore machines).

hs = if 1, hierarchical softmax will be used for model training. If set to 0 (default), and negative is non-zero, negative sampling will be used.

negative = if > 0, negative sampling will be used, the int for negative specifies how many “noise words” should be drawn (usually between 5-20). Default is 5. If set to 0, no negative samping is used.

cbow_mean = if 0, use the sum of the context word vectors. If 1 (default), use the mean. Only applies when cbow is used.

hashfxn = hash function to use to randomly initialize weights, for increased training reproducibility. Default is Python’s rudimentary built in hash function.

iter = number of iterations (epochs) over the corpus. Default is 5.

trim_rule = vocabulary trimming rule, specifies whether certain words should remain in the vocabulary, be trimmed away, or handled using the default (discard if word count < min_count). Can be None (min_count will be used), or a callable that accepts parameters (word, count, min_count) and returns either util.RULE_DISCARD, util.RULE_KEEP or util.RULE_DEFAULT. Note: The rule, if given, is only used prune vocabulary during build_vocab() and is not stored as part of the model.

sorted_vocab = if 1 (default), sort the vocabulary by descending frequency before assigning word indexes.

batch_words = target size (in words) for batches of examples passed to worker threads (and thus cython routines). Default is 10000. (Larger batches will be passed if individual texts are longer than 10000 words, but the standard cython code truncates to that maximum.)

accuracy(questions, restrict_vocab=30000, most_similar=<function most_similar>, use_lowercase=True)

Compute accuracy of the model. questions is a filename where lines are 4-tuples of words, split into sections by ”: SECTION NAME” lines. See questions-words.txt in for an example.

The accuracy is reported (=printed to log and returned as a list) for each section separately, plus there’s one aggregate summary at the end.

Use restrict_vocab to ignore all questions containing a word whose frequency is not in the top-N most frequent words (default top 30,000).

Use use_lowercase to convert all words in questions to thier lowercase form before evaluating the accuracy. It’s useful when assuming the text preprocessing also uses lowercase. (default True).

This method corresponds to the compute-accuracy script of the original C word2vec.

build_vocab(sentences, keep_raw_vocab=False, trim_rule=None, progress_per=10000)

Build vocabulary from a sequence of sentences (can be a once-only generator stream). Each sentence must be a list of unicode strings.


Create a binary Huffman tree using stored vocabulary word counts. Frequent words will have shorter binary codes. Called internally from build_vocab().


Which word from the given list doesn’t go with the others?


>>> trained_model.doesnt_match("breakfast cereal dinner lunch".split())
estimate_memory(vocab_size=None, report=None)

Estimate required memory for a model using current settings and provided vocabulary size.


Build tables and model weights based on final vocabulary settings.


Precompute L2-normalized vectors.

If replace is set, forget the original vectors and only keep the normalized ones = saves lots of memory!

Note that you cannot continue training after doing a replace. The model becomes effectively read-only = you can call most_similar, similarity etc., but not train.

intersect_word2vec_format(fname, binary=False, encoding='utf8', unicode_errors='strict')

Merge the input-hidden weight matrix from the original C word2vec-tool format given, where it intersects with the current vocabulary. (No words are added to the existing vocabulary, but intersecting words adopt the file’s weights, and non-intersecting words are left alone.)

binary is a boolean indicating whether the data is in binary word2vec format.

classmethod load(*args, **kwargs)
classmethod load_word2vec_format(fname, fvocab=None, binary=False, encoding='utf8', unicode_errors='strict')

Load the input-hidden weight matrix from the original C word2vec-tool format.

Note that the information stored in the file is incomplete (the binary tree is missing), so while you can query for word similarity etc., you cannot continue training with a model loaded this way.

binary is a boolean indicating whether the data is in binary word2vec format. norm_only is a boolean indicating whether to only store normalised word2vec vectors in memory. Word counts are read from fvocab filename, if set (this is the file generated by -save-vocab flag of the original C tool).

If you trained the C model using non-utf8 encoding for words, specify that encoding in encoding.

static log_accuracy(section)
make_cum_table(power=0.75, domain=2147483647)

Create a cumulative-distribution table using stored vocabulary word counts for drawing random words in the negative-sampling training routines.

To draw a word index, choose a random integer up to the maximum value in the table (cum_table[-1]), then finding that integer’s sorted insertion point (as if by bisect_left or ndarray.searchsorted()). That insertion point is the drawn index, coming up in proportion equal to the increment at that slot.

Called internally from ‘build_vocab()’.

most_similar(positive=[], negative=[], topn=10, restrict_vocab=None)

Find the top-N most similar words. Positive words contribute positively towards the similarity, negative words negatively.

This method computes cosine similarity between a simple mean of the projection weight vectors of the given words and the vectors for each word in the model. The method corresponds to the word-analogy and distance scripts in the original word2vec implementation.

If topn is False, most_similar returns the vector of similarity scores.

restrict_vocab is an optional integer which limits the range of vectors which are searched for most-similar values. For example, restrict_vocab=10000 would only check the first 10000 word vectors in the vocabulary order. (This may be meaningful if you’ve sorted the vocabulary by descending frequency.)


>>> trained_model.most_similar(positive=['woman', 'king'], negative=['man'])
[('queen', 0.50882536), ...]
most_similar_cosmul(positive=[], negative=[], topn=10)

Find the top-N most similar words, using the multiplicative combination objective proposed by Omer Levy and Yoav Goldberg in [4]. Positive words still contribute positively towards the similarity, negative words negatively, but with less susceptibility to one large distance dominating the calculation.

In the common analogy-solving case, of two positive and one negative examples, this method is equivalent to the “3CosMul” objective (equation (4)) of Levy and Goldberg.

Additional positive or negative examples contribute to the numerator or denominator, respectively – a potentially sensible but untested extension of the method. (With a single positive example, rankings will be the same as in the default most_similar.)


>>> trained_model.most_similar_cosmul(positive=['baghdad', 'england'], negative=['london'])
[(u'iraq', 0.8488819003105164), ...]
[4]Omer Levy and Yoav Goldberg. Linguistic Regularities in Sparse and Explicit Word Representations, 2014.
n_similarity(ws1, ws2)

Compute cosine similarity between two sets of words.


>>> trained_model.n_similarity(['sushi', 'shop'], ['japanese', 'restaurant'])

>>> trained_model.n_similarity(['restaurant', 'japanese'], ['japanese', 'restaurant'])

>>> trained_model.n_similarity(['sushi'], ['restaurant']) == trained_model.similarity('sushi', 'restaurant')

Borrow shareable pre-built structures (like vocab) from the other_model. Useful if testing multiple models in parallel on the same corpus.


Reset all projection weights to an initial (untrained) state, but keep the existing vocabulary.

save(*args, **kwargs)

Save the object to file (also see load).

fname_or_handle is either a string specifying the file name to save to, or an open file-like object which can be written to. If the object is a file handle, no special array handling will be performed; all attributes will be saved to the same file.

If separately is None, automatically detect large numpy/scipy.sparse arrays in the object being stored, and store them into separate files. This avoids pickle memory errors and allows mmap’ing large arrays back on load efficiently.

You can also set separately manually, in which case it must be a list of attribute names to be stored in separate files. The automatic check is not performed in this case.

ignore is a set of attribute names to not serialize (file handles, caches etc). On subsequent load() these attributes will be set to None.

pickle_protocol defaults to 2 so the pickled object can be imported in both Python 2 and 3.

save_word2vec_format(fname, fvocab=None, binary=False)

Store the input-hidden weight matrix in the same format used by the original C word2vec-tool, for compatibility.

fname is the file used to save the vectors in fvocab is an optional file used to save the vocabulary binary is an optional boolean indicating whether the data is to be saved in binary word2vec format (default: False)
scale_vocab(min_count=None, sample=None, dry_run=False, keep_raw_vocab=False, trim_rule=None)

Apply vocabulary settings for min_count (discarding less-frequent words) and sample (controlling the downsampling of more-frequent words).

Calling with dry_run=True will only simulate the provided settings and report the size of the retained vocabulary, effective corpus length, and estimated memory requirements. Results are both printed via logging and returned as a dict.

Delete the raw vocabulary after the scaling is done to free up RAM, unless keep_raw_vocab is set.

scan_vocab(sentences, progress_per=10000, trim_rule=None)

Do an initial scan of all words appearing in sentences.

score(sentences, total_sentences=1000000, chunksize=100, queue_factor=2, report_delay=1)

Score the log probability for a sequence of sentences (can be a once-only generator stream). Each sentence must be a list of unicode strings. This does not change the fitted model in any way (see Word2Vec.train() for that).

We have currently only implemented score for the hierarchical softmax scheme, so you need to have run word2vec with hs=1 and negative=0 for this to work.

Note that you should specify total_sentences; we’ll run into problems if you ask to score more than this number of sentences but it is inefficient to set the value too high.

See the article by [taddy] and the gensim demo at [deepir] for examples of how to use such scores in document classification.

[taddy]Taddy, Matt. Document Classification by Inversion of Distributed Language Representations, in Proceedings of the 2015 Conference of the Association of Computational Linguistics.

Create one ‘random’ vector (but deterministic by seed_string)

similar_by_vector(vector, topn=10, restrict_vocab=None)

Find the top-N most similar words by vector.

If topn is False, similar_by_vector returns the vector of similarity scores.

restrict_vocab is an optional integer which limits the range of vectors which are searched for most-similar values. For example, restrict_vocab=10000 would only check the first 10000 word vectors in the vocabulary order. (This may be meaningful if you’ve sorted the vocabulary by descending frequency.)


>>> trained_model.similar_by_vector([1,2])
[('survey', 0.9942699074745178), ...]
similar_by_word(word, topn=10, restrict_vocab=None)

Find the top-N most similar words.

If topn is False, similar_by_word returns the vector of similarity scores.

restrict_vocab is an optional integer which limits the range of vectors which are searched for most-similar values. For example, restrict_vocab=10000 would only check the first 10000 word vectors in the vocabulary order. (This may be meaningful if you’ve sorted the vocabulary by descending frequency.)


>>> trained_model.similar_by_word('graph')
[('user', 0.9999163150787354), ...]
similarity(w1, w2)

Compute cosine similarity between two words.


>>> trained_model.similarity('woman', 'man')

>>> trained_model.similarity('woman', 'woman')

Sort the vocabulary so the most frequent words have the lowest indexes.

train(sentences, total_words=None, word_count=0, total_examples=None, queue_factor=2, report_delay=1.0)

Update the model’s neural weights from a sequence of sentences (can be a once-only generator stream). For Word2Vec, each sentence must be a list of unicode strings. (Subclasses may accept other examples.)

To support linear learning-rate decay from (initial) alpha to min_alpha, either total_examples (count of sentences) or total_words (count of raw words in sentences) should be provided, unless the sentences are the same as those that were used to initially build the vocabulary.

wmdistance(document1, document2, WCD=False, RWMD=False)

Compute the Word Mover’s Distance between two documents. When using this code, please consider citing the following papers: * Ofir Pele and Michael Werman, “A linear time histogram metric for improved SIFT matching”. * Ofir Pele and Michael Werman, “Fast and robust earth mover’s distances”. * Matt Kusner et al. “From Word Embeddings To Document Distances”.

Note that if one of the documents have no words that exist in the Word2Vec vocab, float(‘inf’) (i.e. infinity) will be returned.

This method only works if pyemd is installed (can be installed via pip, but requires a C compiler).

Example: > # Train word2vec model. > model = Word2Vec(sentences)

> # Some sentences to test. > sentence_obama = ‘Obama speaks to the media in Illinois’.lower().split() > sentence_president = ‘The president greets the press in Chicago’.lower().split()

> # Remove their stopwords. > from nltk.corpus import stopwords > stopwords = nltk.corpus.stopwords.words(‘english’) > sentence_obama = [w for w in sentence_obama if w not in stopwords] > sentence_president = [w for w in sentence_president if w not in stopwords]

> # Compute WMD. > distance = model.wmdistance(sentence_obama, sentence_president)

gensim.models.word2vec.score_cbow_pair(model, word, word2_indices, l1)
gensim.models.word2vec.score_sg_pair(model, word, word2)
gensim.models.word2vec.train_cbow_pair(model, word, input_word_indices, l1, alpha, learn_vectors=True, learn_hidden=True)
gensim.models.word2vec.train_sg_pair(model, word, context_index, alpha, learn_vectors=True, learn_hidden=True, context_vectors=None, context_locks=None)