sense2vec: Use NLP to go beyond vanilla word2vec

❗️❗️❗️ This fork focuses on making s2v work (with spaCy 2.x) in particular to train new models (for non-English languages). Please do not use pip to install. Read more under installation.

❗️❗️❗️ We also plan to release some models in different languages that have NER+Tagging support in spaCy. For requests, open an issue and clearly mention where to find the NER+Tragging model for the language you want.

sense2vec (Trask et. al, 2015) is a nice twist on word2vec that lets you learn more interesting, detailed and context-sensitive word vectors. For an interactive example of the technology, see our sense2vec demo that lets you explore semantic similarities across all Reddit comments of 2015.

This library is a simple Python/Cython implementation for loading and querying sense2vec models. While it's best used in combination with spaCy, the sense2vec library itself is very lightweight and can also be used as a standalone module. See below for usage details.

🦆 Version 1.0 alpha out now! Read the release notes here.

Usage Examples

Usage with spaCy

import spacy
from sense2vec import Sense2VecComponent

nlp = spacy.load('en')
s2v = Sense2VecComponent('/path/to/reddit_vectors-1.1.0')
nlp.add_pipe(s2v)

doc = nlp(u"A sentence about natural language processing.")
assert doc[3].text == u'natural language processing'
freq = doc[3]._.s2v_freq
vector = doc[3]._.s2v_vec
most_similar = doc[3]._.s2v_most_similar(3)
# [(('natural language processing', 'NOUN'), 1.0),
#  (('machine learning', 'NOUN'), 0.8986966609954834),
#  (('computer vision', 'NOUN'), 0.8636297583580017)]

Standalone usage without spaCy

import sense2vec

s2v = sense2vec.load('/path/to/reddit_vectors-1.1.0')
query = u'natural_language_processing|NOUN'
assert query in s2v
freq, vector = s2v[query]
words, scores = s2v.most_similar(vector, 3)
most_similar = list(zip(words, scores))
# [('natural_language_processing|NOUN', 1.0),
#  ('machine_learning|NOUN', 0.8986966609954834),
#  ('computer_vision|NOUN', 0.8636297583580017)]

Installation & Setup

Operating system	macOS / OS X, Linux, Windows (Cygwin, MinGW, Visual Studio)
Python version	CPython 2.7, 3.4+. Only 64 bit.
Package managers	pip (source packages only)

❗️❗️❗️ For this fork to work, you need to clone it first and build and install locally. If you already have s2v, remove it first (tested under Python 3 only):

pip uninstall sense2vec
python setup.py build
python setup.py install

sense2vec releases are available as source packages on pip:

pip install sense2vec==1.0.0a0

The Reddit vectors model is attached to the latest release. To load it in, download the .tar.gz archive, unpack it and point sense2vec.load to the extracted data directory:

import sense2vec
s2v = sense2vec.load('/path/to/reddit_vectors-1.1.0')

Usage

Usage with spaCy v2.x

The easiest way to use the library and vectors is to plug it into your spaCy pipeline. Note that sense2vec doesn't depend on spaCy, so you'll have to install it separately and download the English model.

pip install -U spacy
python -m spacy download en

The sense2vec package exposes a Sense2VecComponent, which can be initialised with the data path and added to your spaCy pipeline as a custom pipeline component. By default, components are added to the end of the pipeline, which is the recommended position for this component, since it needs access to the dependency parse and, if available, named entities.

import spacy
from sense2vec import Sense2VecComponent

nlp = spacy.load('en')
s2v = Sense2VecComponent('/path/to/reddit_vectors-1.1.0')
nlp.add_pipe(s2v)

The pipeline component will merge noun phrases and entities according to the same schema used when training the sense2vec models (e.g. noun chunks without determiners like "the"). This ensures that you'll be able to retrieve meaningful vectors for phrases in your text. The component will also add serveral extension attributes and methods to spaCy's Token and Span objects that let you retrieve vectors and frequencies, as well as most similar terms.

doc = nlp(u"A sentence about natural language processing.")
assert doc[3].text == u'natural language processing'
assert doc[3]._.in_s2v
freq = doc[3]._.s2v_freq
vector = doc[3]._.s2v_vec
most_similar = doc[3]._.s2v_most_similar(10)

For entities, the entity labels are used as the "sense" (instead of the token's part-of-speech tag):

doc = nlp(u"A sentence about Facebook and Google.")
for ent in doc.ents:
    assert ent._.in_s2v
    most_similar = ent._.s2v_most_similar(3)

Available attributes

The following attributes are available via the ._ property – for example token._.in_s2v:

Name	Attribute Type	Type	Description
in_s2v	property	bool	Whether a key exists in the vector map.
s2v_freq	property	int	The frequency of the given key.
s2v_vec	property	ndarray[float32]	The vector of the given key.
s2v_most_similar	method	list	Get the n most similar terms. Returns a list of ((word, sense), score) tuples.

A note on span attributes: Under the hood, entities in doc.ents are Span objects. This is why the pipeline component also adds attributes and methods to spans and not just tokens. However, it's not recommended to use the sense2vec attributes on arbitrary slices of the document, since the model likely won't have a key for the respective text. Span objects also don't have a part-of-speech tag, so if no entity label is present, the "sense" defaults to the root's part-of-speech tag.

Standalone usage

To use only the sense2vec library, you can import the package and then call its load() method to load in the vectors.

import sense2vec
s2v = sense2vec.load('/path/to/reddit_vectors-1.1.0')

sense2vec.load returns an instance of the VectorMap class, which you can interact with via the following methods.

⚠️ Important note: When interacting with the VectorMap directly, the keys need to follow the scheme of phrase_text|SENSE (note the _ instead of spaces and the | before the tag or label) – for example, machine_learning|NOUN. Also note that the underlying vector table is case-sensitive.

VectorMap.__len__

The total number of entries in the map.

Argument	Type	Description
RETURNS	int	The number of entries in the map.

s2v = sense2vec.load('/path/to/reddit_vectors-1.1.0')
assert len(s2v) == 1195261

VectorMap.__contains__

Check whether the VectorMap has a given key. Keys consist of the word string, a pipe and the "sense", i.e. the part-of-speech tag or entity label. For example: 'duck|NOUN' or 'duck|VERB'. See the section on "Senses" below for more details. Also note that the underlying vector table is case-sensitive.

Argument	Type	Description
string	unicode	The key to check.
RETURNS	bool	Whether the key is part of the map.

assert u'duck|NOUN' in s2v
assert u'duck|VERB' in s2v
assert u'dkdksl|VERB' not in s2v

VectorMap.__getitem__

Retrieve a (frequency, vector) tuple from the vector map. The frequency is an integer, the vector a numpy.ndarray(dtype='float32'). If the key is not found, a KeyError is raised.

Argument	Type	Description
string	unicode	The key to retrieve the frequency and vector for.
RETURNS	tuple	The (frequency, vector) tuple.

freq, vector = s2v[u'duck|NOUN']

VectorMap.__setitem__

Assign a (frequency, vector) tuple to the vector map. The frequency should be an integer, the vector a numpy.ndarray(dtype='float32').

Argument	Type	Description
key	unicode	The key to assign the frequency and vector to.
value	tuple	The (frequency, vector) tuple to assign.

freq, vector = s2v[u'avocado|NOUN']
s2v[u'🥑|NOUN'] = (freq, vector)

VectorMap.__iter__, VectorMap.keys

Iterate over the keys in the map, in order of insertion.

Argument	Type	Description
YIELDS	unicode	The keys in the map.

VectorMap.values

Iterate over the values in the map, in order of insertion and yield (frequency, vector) tuples from the vector map. The frequency is an integer, the vector a numpy.ndarray(dtype='float32')

Argument	Type	Description
YIELDS	tuple	The values in the map.

VectorMap.items

Iterate over the items in the map, in order of insertion and yield (key, (frequency, vector)) tuples from the vector map. The frequency is an integer, the vector a numpy.ndarray(dtype='float32')

Argument	Type	Description
YIELDS	tuple	The items in the map.

VectorMap.most_similar

Find the keys of the n most similar entries, given a vector. Note that the most similar entry with a score of 1.0 will be the key of the query vector itself.

Argument	Type	Description
vector	numpy.ndarray(dtype='float32')	The vector to compare to.
n	int	The number of entries to return. Defaults to 10.
RETURNS	tuple	A (words, scores) tuple.

freq, vector = s2v[u'avocado|NOUN']
words, scores = s2v.most_similar(vector, n=3)
for word, score in zip(words, scores):
    print(word, score)
# avocado|NOUN 1.0
# avacado|NOUN 0.970944344997406
# spinach|NOUN 0.962776780128479

VectorMap.save

Serialize the model to a directory. This will export three files to the output directory: a strings.json containing the keys in insertion order, a freqs.json containing the frequencies and a vectors.bin containing the vectors.

Argument	Type	Description
data_dir	unicode	The path to the output directory.

VectorMap.load

Load a model from a directory. Expects three files in the directory (see VectorMap.save for details).

Argument	Type	Description
data_dir	unicode	The path to load the model from.

Senses

The pre-trained Reddit vectors support the following "senses", either part-of-speech tags or entity labels. For more details, see spaCy's annotation scheme overview.

Tag	Description	Examples
ADJ	adjective	big, old, green
ADP	adposition	in, to, during
ADV	adverb	very, tomorrow, down, where
AUX	auxiliary	is, has (done), will (do)
CONJ	conjunction	and, or, but
DET	determiner	a, an, the
INTJ	interjection	psst, ouch, bravo, hello
NOUN	noun	girl, cat, tree, air, beauty
NUM	numeral	1, 2017, one, seventy-seven, MMXIV
PART	particle	's, not
PRON	pronoun	I, you, he, she, myself, somebody
PROPN	proper noun	Mary, John, London, NATO, HBO
PUNCT	punctuation	, ? ( )
SCONJ	subordinating conjunction	if, while, that
SYM	symbol	$, %, =, :), 😝
VERB	verb	run, runs, running, eat, ate, eating

Entity Label	Description
PERSON	People, including fictional.
NORP	Nationalities or religious or political groups.
FACILITY	Buildings, airports, highways, bridges, etc.
ORG	Companies, agencies, institutions, etc.
GPE	Countries, cities, states.
LOC	Non-GPE locations, mountain ranges, bodies of water.
PRODUCT	Objects, vehicles, foods, etc. (Not services.)
EVENT	Named hurricanes, battles, wars, sports events, etc.
WORK_OF_ART	Titles of books, songs, etc.
LANGUAGE	Any named language.

Training a sense2vec model

Before training a model you need to have a large enough dataset and preprocess it using bin/preprocess.py. For models we release, we use a combination of a full Wikipedia articles dump with a large amount of tweets that Twitter classified to be in the same language.

The bin/train.py script can be used to train new models. By default the training is done using w2v from Gensim but if you want to use another algorithm like GloVe or FastText you can modify train.py (FastText example is commented out). We experimented with both w2v and FastText and found that FastText quickly focuses too heavily on common POS-tags and hence is most likely not your preferred choice. We also found that using a large number of epochs (>100) drastically overfits your vectors. We use the default of 5 everywhere.

Full spaCy pipeline models

We like completeness. Hence we want to have the NER, tagging, parsing, word vectors and sense vectors all combined into one spaCy pipeline. One way to do this is as follows, where we load a default spaCy model containing NER, tagging and parsing, add in word vectors downloaded from FastText (or any other source, see spaCy CLI on how to do this (init first, then package)) and our own trained sense vectors.

import spacy
from sense2vec import Sense2VecComponent

nlp = spacy.load('nl_core_news_sm') # Default spaCy model with NER, tagging, parsing
nlp.vocab.vectors.from_disk('[LOCATION_TO_WORD_VECTOR_MODEL]') # Replace here with the word vector model you produced after spacy package
s2v = Sense2VecComponent('[LOCATION_TO_S2V_MODEL]', 300) # Replace with the pre-trained s2v model
nlp.add_pipe(s2v) # Add s2v to pipeline
doc = nlp("Ik ben in New York") # I am in New York
doc[3]._.s2v_most_similar(3) # Most similar to New_York|LOC

Pre-trained models:

NL - https://drive.google.com/open?id=1TC_XJwDHNTD5ir68THOj6aaX44M68kCR