Wikipedia2Vec is a tool used for obtaining embeddings (vector representations) of words and entities from Wikipedia. It is developed and maintained by Studio Ousia.
This tool enables you to learn embeddings that map words and entities into a unified continuous vector space. The embeddings can be used as word embeddings, entity embeddings, and the unified embeddings of words and entities. They are used in the state-of-the-art models of various tasks such as entity linking, named entity recognition, entity relatedness, and question answering.
The embeddings can be easily trained from a publicly available Wikipedia dump. The code is implemented in Python, and optimized using Cython and BLAS.
Wikipedia2Vec is based on the Word2vec's skip-gram model that learns to predict neighboring words given each word in corpora. We extend the skip-gram model by adding the following two submodels:
- The KB link graph model that learns to estimate neighboring entities given an entity in the link graph of Wikipedia entities.
- The anchor context model that learns to predict neighboring words given an entity by using an anchor link that points to the entity and its neighboring words.
By jointly optimizing the skip-gram model and these two submodels, our model simultaneously learns the embedding of words and entities from Wikipedia. For further details, please refer to our paper: Joint Learning of the Embedding of Words and Entities for Named Entity Disambiguation.
(coming soon)
If you want to train embeddings on your machine, it is highly recommended to install a BLAS library before installing this tool. We recommend using OpenBLAS or Intel Math Kernel Library.
Wikipedia2Vec can be installed from PyPI:
% pip install wikipedia2vec
The command installs the following required Python libraries: click, jieba, joblib, lmdb, marisa-trie, mwparserfromhell, numpy, scipy, six, and tqdm.
To process Japanese Wikipedia dumps, it is also required to install MeCab and its Python binding.
Currently, this software is tested only on Linux.
First, you need to download a source Wikipedia dump file (e.g., enwiki-latest-pages-articles.xml.bz2) from Wikimedia Downloads. The English dump file can be obtained by running the following command.
% wget https://dumps.wikimedia.org/enwiki/latest/enwiki-latest-pages-articles.xml.bz2
Note that you do not need to decompress the dump file.
Then, the embeddings can be trained from a Wikipedia dump using the train command.
% wikipedia2vec train DUMP_FILE OUT_FILE
Arguments:
- DUMP_FILE: The Wikipedia dump file
- OUT_FILE: The output file
Options:
- --dim-size: The number of dimensions of the embeddings (default: 100)
- --window: The maximum distance between the target item (word or entity) and the context word to be predicted (default: 5)
- --iteration: The number of iterations for Wikipedia pages (default: 3)
- --negative: The number of negative samples (default: 5)
- --lowercase/--no-lowercase: Whether to lowercase words (default: True)
- --min-word-count: A word is ignored if the total frequency of the word is less than this value (default: 10)
- --min-entity-count: An entity is ignored if the total frequency of the entity appearing as the referent of an anchor link is less than this value (default: 5)
- --min-paragraph-len: A paragraph is ignored if its length is shorter than this value (default: 5)
- --category/--no-category: Whether to include Wikipedia categories in the dictionary (default:False)
- --link-graph/--no-link-graph: Whether to learn from the Wikipedia link graph (default: True)
- --entities-per-page: For processing each page, the specified number of randomly chosen entities are used to predict their neighboring entities in the link graph (default: 5)
- --init-alpha: The initial learning rate (default: 0.025)
- --min-alpha: The minimum learning rate (default: 0.0001)
- --sample: The parameter that controls the downsampling of frequent words (default: 1e-4)
- --word-neg-power: Negative sampling of words is performed based on the probability proportional to the frequency raised to the power specified by this option (default: 0.75)
- --entity-neg-power: Negative sampling of entities is performed based on the probability proportional to the frequency raised to the power specified by this option (default: 0)
The train command internally calls the four commands described below (namely, build_dump_db, build_dictionary, build_link_graph, and train_embedding).
The build_dump_db command creates a database that contains Wikipedia pages each of which consists of texts and anchor links in it. The size of the database based on an English Wikipedia dump is approximately 15GB.
% wikipedia2vec build_dump_db DUMP_FILE OUT_FILE
Arguments:
- DUMP_FILE: The Wikipedia dump file
- OUT_FILE: The output file
The build_dictionary command builds a dictionary of words and entities.
% wikipedia2vec build_dictionary DUMP_DB_FILE OUT_FILE
Arguments:
- DUMP_DB_FILE: The database file generated using the build_dump_db command
- OUT_FILE: The output file
Options:
- --lowercase/--no-lowercase: Whether to lowercase words (default: True)
- --min-word-count: A word is ignored if the total frequency of the word is less than this value (default: 10)
- --min-entity-count: An entity is ignored if the total frequency of the entity appearing as the referent of an anchor link is less than this value (default: 5)
- --min-paragraph-len: A paragraph is ignored if its length is shorter than this value (default: 5)
- --category/--no-category: Whether to include Wikipedia categories in the dictionary (default:False)
The build_link_graph command generates a sparse matrix representing the link structure between Wikipedia entities.
% wikipedia2vec build_link_graph DUMP_DB_FILE DIC_FILE OUT_FILE
Arguments:
- DUMP_DB_FILE: The database file generated using the build_dump_db command
- DIC_FILE: The dictionary file generated by the build_dictionary command
- OUT_FILE: The output file
There is no option in this command.
The train_embedding command runs the training of the embeddings.
% wikipedia2vec train_embedding DUMP_DB_FILE DIC_FILE OUT_FILE
Arguments:
- DUMP_DB_FILE: The database file generated using the build_dump_db command
- DIC_FILE: The dictionary file generated by the build_dictionary command
- OUT_FILE: The output file
Options:
- --link-graph: The link graph file generated using the build_link_graph command
- --dim-size: The number of dimensions of the embeddings (default: 100)
- --window: The maximum distance between the target item (word or entity) and the context word to be predicted (default: 5)
- --iteration: The number of iterations for Wikipedia pages (default: 3)
- --negative: The number of negative samples (default: 5)
- --word-neg-power: Negative sampling of words is performed based on the probability proportional to the frequency raised to the power specified by this option (default: 0.75)
- --entity-neg-power: Negative sampling of entities is performed based on the probability proportional to the frequency raised to the power specified by this option (default: 0)
- --entities-per-page: For processing each page, the specified number of randomly chosen entities are used to predict their neighboring entities in the link graph (default: 10)
- --init-alpha: The initial learning rate (default: 0.025)
- --min-alpha: The minimum learning rate (default: 0.0001)
- --sample: The parameter that controls the downsampling of frequent words (default: 1e-4)
save_text outputs a model in a text format.
% wikipedia2vec save_text MODEL_FILE OUT_FILE
Arguments:
- MODEL_FILE: The model file generated by the train_embedding command
- OUT_FILE: The output file
Options:
- --out-format: The output format. Possible values are default, word2vec, and glove. If word2vec and glove are specified, the format adopted by Word2Vec and GloVe are used, respectively.
>>> from wikipedia2vec import Wikipedia2Vec
>>> wiki2vec = Wikipedia2Vec.load(MODEL_FILE)
>>> wiki2vec.get_word_vector(u'the')
memmap([ 0.01617998, -0.03325786, -0.01397999, -0.00150471, 0.03237337,
...
-0.04226106, -0.19677088, -0.31087297, 0.1071524 , -0.09824426], dtype=float32)
>>> wiki2vec.get_entity_vector(u'Scarlett Johansson')
memmap([-0.19793572, 0.30861306, 0.29620451, -0.01193621, 0.18228433,
...
0.04986198, 0.24383858, -0.01466644, 0.10835337, -0.0697331 ], dtype=float32)
>>> wiki2vec.most_similar(wiki2vec.get_word(u'yoda'), 5)
[(<Word yoda>, 1.0),
(<Entity Yoda>, 0.84333622),
(<Word darth>, 0.73328167),
(<Word kenobi>, 0.7328127),
(<Word jedi>, 0.7223742)]
>>> wiki2vec.most_similar(wiki2vec.get_entity(u'Scarlett Johansson'), 5)
[(<Entity Scarlett Johansson>, 1.0),
(<Entity Natalie Portman>, 0.75090045),
(<Entity Eva Mendes>, 0.73651594),
(<Entity Emma Stone>, 0.72868186),
(<Entity Cameron Diaz>, 0.72390842)]
If you use Wikipedia2Vec in a scientific publication, please cite the following paper:
@InProceedings{yamada-EtAl:2016:CoNLL,
author = {Yamada, Ikuya and Shindo, Hiroyuki and Takeda, Hideaki and Takefuji, Yoshiyasu},
title = {Joint Learning of the Embedding of Words and Entities for Named Entity Disambiguation},
booktitle = {Proceedings of The 20th SIGNLL Conference on Computational Natural Language Learning},
month = {August},
year = {2016},
address = {Berlin, Germany},
pages = {250--259},
publisher = {Association for Computational Linguistics}
}