wikirec 0.1.1.9

NLP recommendation engine based on Wikipedia data

---

NLP recommendation engine based on Wikipedia data

Jump to: Data • Methods • Recommendations • Comparative Results • To-Do

wikirec is a framework that allows users to parse Wikipedia in any language for entries of a given type and then seamlessly generate recommendations based on unsupervised natural language processing. The gaol is for wikirec to both refine and deploy models that provide accurate content recommendations based solely on open-source data. See the documentation for a full outline of the package including models and data preparation.

Installation via PyPi

wikirec can be downloaded from pypi via pip or sourced directly from this repository:

pip install wikirec

git clone https://github.com/andrewtavis/wikirec.git
cd wikirec
python setup.py install

import wikirec

Data ↩

wikirec.data_utils allows a user to download Wikipedia articles of a given topic including movies, TV shows, books, music, and countless other classes of information. These texts then serve as the basis to recommend similar content given an input of what the user is interested in.

Article topics are derived from infobox types found on Wikipedia articles. The article on infoboxes (and its translations) contains all the allowed arguments to subset the data by. Simply passing "Infobox chosen_type" to the topics argument of data_utils.parse_to_ndjson() in the following example will subset all Wikipedia articles for the given type. Lists can also be passed if more than one topic is desired. For the English Wikipedia, wikirec also provides concise arguments for data that commonly serve as recommendation inputs including: books, songs, albums, movies, tv_series, video_games, as well as various categories of people such as athletes, musicians and authors (see data_utils.input_conversion_dict()).

Downloading and parsing Wikipedia for the needed data is as simple as:

from wikirec import data_utils

# Downloads the most recent stable bz2 compressed English Wikipedia dump
files = data_utils.download_wiki(language="en", target_dir="enwiki_dump")

# Produces an ndjson of all book articles on Wikipedia
data_utils.parse_to_ndjson(
    topics="books",  # ["books", "short_stories", "plays"]
    output_path="enwiki_books.ndjson",
    input_dir="enwiki_dump",
    limit=None,  # articles per file to find
    multicore=True,
    verbose=True,
)

The examples directory has a compressed copy of enwiki_books.ndjson for testing purposes.

wikirec.data_utils also provides a standardized multilingual cleaning process for the loaded articles. See wikirec.languages for a full breakdown of what is available for each language. Generating a clean text corpus is achieved through the following:

import json

with open("enwiki_books.ndjson", "r") as f:
    books = [json.loads(l) for l in f]

titles = [b[0] for b in books]
texts = [b[1] for b in books]

text_corpus, selected_idxs = data_utils.clean(
    texts=texts,
    language="en",
    min_token_freq=5,  # 0 for Bert
    min_token_len=3,  # 0 for Bert
    min_tokens=50,
    max_token_index=-1,
    remove_stopwords=True,  # False for Bert
    verbose=True,
)

selected_titles = [titles[i] for i in selected_idxs]

Methods ↩

Recommendations in wikirec are generated from similarity matrices derived from trained model embeddings. Implemented NLP modeling methods within wikirec.model include:

BERT

Bidirectional Encoder Representations from Transformers derives representations of words based on NLP models ran over open source Wikipedia data. These representations are leveraged to derive article similarities that are then used to deliver recommendations.

wikirec uses sentence-transformers pretrained models. See their GitHub and documentation for the available models.

from wikirec import model

# Remove n-grams for BERT training
corpus_no_ngrams = [
    " ".join([t for t in text.split(" ") if "_" not in t]) for text in text_corpus
]

# We can pass kwargs for sentence_transformers.SentenceTransformer.encode
bert_embeddings = model.gen_embeddings(
        method="bert",
        corpus=corpus_no_ngrams,
        bert_st_model="xlm-r-bert-base-nli-stsb-mean-tokens",
        batch_size=32,
)

Doc2vec

A generalization of Word2vec, Doc2vec is an NLP algorithm for deriving vector representations of documents from contextual word interrelations. These representations are then used as a baseline for recommendations.

from wikirec import model

# We can pass kwargs for gensim.models.doc2vec.Doc2Vec
d2v_embeddings = model.gen_embeddings(
        method="doc2vec",
        corpus=text_corpus,
        vector_size=100,
        epochs=10,
        alpha=0.025,
)

LDA

Latent Dirichlet Allocation is a generative statistical model that allows sets of observations to be explained by unobserved groups that explain why some parts of the data are similar. In the case of wikirec, Wikipedia articles are posited to be a mixture of a given number of topics, and the presence of each word in a text body comes from its relation to these derived topics. These topic-word relations are then used to determine article similarities and then make recommendations.

from wikirec import model

# We can pass kwargs for gensim.models.ldamulticore.LdaMulticore
lda_embeddings = model.gen_embeddings(
        method="lda",
        corpus=text_corpus,  # automatically tokenized for LDA
        num_topics=50,
        passes=10,
        decay=0.5,
)

TFIDF

Term Frequency Inverse Document Frequency is a numerical statistic that is intended to reflect how important a word is to a document in a collection or corpus. In case of wikirec, word importances are combined and compared to derive article similarities and thus provide recommendations.

from wikirec import model

# We can pass kwargs for sklearn.feature_extraction.text.TfidfVectorizer
tfidf_embeddings = model.gen_embeddings(
        method="tfidf",
        corpus=text_corpus,
        max_features=None,
        norm="l2",
)

Recommendations ↩

After embeddings have been generated we can then create matrices that represent article-article cosine or euclidean similarities. These can then be sorted and selected from, with the recommendation process being as simple as the following:

from wikirec import model

sim_matrix = model.gen_sim_matrix(
    method="chosen_method",
    metric="cosine",  # euclidean
    embeddings=method_embeddings,
)

recs = model.recommend(
    inputs="title_or_list_of_titles",
    titles=selected_titles,
    sim_matrix=sim_matrix,
    metric="cosine",  # euclidean
    n=10,
)

Comparative Results ↩

TFIDF generally outperformed all other methods in terms of providing what the user would expect, with the results being all the more striking considering its runtime is by far the shortest. The other strong performing model is BERT, as it does the best job of providing novel but sensible recommendations. LDA with the second shortest runtime provides novel recommendations along with what is expected, but recommends things that seem out of place more often than BERT. Doc2vec performs very poorly in that most results are nonsense, and it further takes the longest to train.

See examples/rec_books and examples/rec_movies for fully detailed usage with model comparisons, or open these notebooks in Google Colab to experiment yourself.

A sample of TFIDF and BERT book recommendations using cosine similarity follows:

--TFIDF--
Harry Potter and the Philosopher's Stone recommendations:
[['Harry Potter and the Chamber of Secrets', 0.5974588223913879],
 ['Harry Potter and the Deathly Hallows', 0.5803045645372675],
 ['Harry Potter and the Goblet of Fire', 0.5752151957878091],
 ['Harry Potter and the Half-Blood Prince', 0.5673108963392828],
 ['Harry Potter and the Order of the Phoenix', 0.5662440277414937],
 ['The Magical Worlds of Harry Potter', 0.5098747039144682],
 ['Harry Potter and the Methods of Rationality', 0.5016950079654786],
 ['Harry Potter and the Prisoner of Azkaban', 0.4865186451505909],
 ['Fantastic Beasts and Where to Find Them', 0.4801163347125484],
 ['The Casual Vacancy', 0.44319508498475246]]

The Hobbit recommendations:
[['The History of The Hobbit', 0.7744692537347045],
 ['The Annotated Hobbit', 0.6474663216496771],
 ['Mr. Bliss', 0.5774314075304691],
 ['The Lord of the Rings', 0.5626569367072154],
 ['The Road to Middle-Earth', 0.5386365684368313],
 ['The Marvellous Land of Snergs', 0.5165174723722297],
 ['Tolkien: Maker of Middle-earth', 0.5062523572124091],
 ['The Letters of J. R. R. Tolkien', 0.489393850451095],
 ['The Tolkien Reader', 0.4862696945481724],
 ['J. R. R. Tolkien: A Biography', 0.4813258277958349]]

Harry Potter and the Philosopher's Stone and The Hobbit recommendations:
[['The History of The Hobbit', 0.4144937936077629],
 ['Harry Potter and the Chamber of Secrets', 0.34888387038976304],
 ['The Lord of the Rings', 0.3461664662907625],
 ['The Annotated Hobbit', 0.3431651523791515],
 ['Harry Potter and the Deathly Hallows', 0.3336208844683567],
 ['Harry Potter and the Goblet of Fire', 0.3323377108209634],
 ['Harry Potter and the Half-Blood Prince', 0.32972615751499673],
 ['Mr. Bliss', 0.3219122094772891],
 ['Harry Potter and the Order of the Phoenix', 0.3160426316664049],
 ['The Magical Worlds of Harry Potter', 0.30770960167033506]]

 --BERT--
 Harry Potter and the Philosopher's Stone recommendations:
[['Harry Potter and the Prisoner of Azkaban', 0.8625375],
 ['Harry Potter and the Chamber of Secrets', 0.8557441],
 ['Harry Potter and the Half-Blood Prince', 0.8430752],
 ['Harry Potter and the Goblet of Fire', 0.8258302],
 ['The Magical Worlds of Harry Potter', 0.82496],
 ['A Bad Spell in Yurt', 0.82023925],
 ['Harry Potter and the Order of the Phoenix', 0.80546284],
 ['So You Want to Be a Wizard', 0.803981],
 ['The Weirdstone of Brisingamen', 0.8035261],
 ['Harry Potter and the Cursed Child', 0.79987496]]

 The Hobbit recommendations:
[['The Lord of the Rings', 0.8724792],
 ['Beast', 0.8283818],
 ['The Children of Húrin', 0.8261733],
 ['The Foundling and Other Tales of Prydain', 0.82471454],
 ['The Black Cauldron', 0.82060313],
 ['El Deafo', 0.8167627],
 ['The Little Grey Men', 0.8116319],
 ['The Woggle-Bug Book', 0.8109094],
 ['The Amazing Maurice and His Educated Rodents', 0.8089799],
 ['Dark Lord of Derkholm', 0.8068354]]

 Harry Potter and the Philosopher's Stone and The Hobbit recommendations:
[['The Weirdstone of Brisingamen', 0.79162943],
 ['Harry Potter and the Prisoner of Azkaban', 0.7681779],
 ['A Wizard of Earthsea', 0.7566709],
 ["The Magician's Nephew", 0.75540984],
 ["Merlin's Wood", 0.7530513],
 ['Harry Potter and the Half-Blood Prince', 0.7483348],
 ['Charmed Life', 0.74817574],
 ['The Borrowers Avenged', 0.7475477],
 ["The Inquisitor's Tale", 0.74703705],
 ['The Ghost of Thomas Kempe', 0.74537575]]

Better results can further be achieved by combining the above two methods:

tfidf_weight = 0.35
bert_weight = 1.0 - tfidf_weight
bert_tfidf_sim_matrix = tfidf_weight * tfidf_sim_matrix + bert_weight * bert_sim_matrix

--Weighted BERT and TFIDF--
 Harry Potter and the Philosopher's Stone recommendations:
[['Harry Potter and the Chamber of Secrets', 0.7653442323224594],
 ['Harry Potter and the Half-Blood Prince', 0.7465576592959889],
 ['Harry Potter and the Goblet of Fire', 0.7381149146065132],
 ['Harry Potter and the Prisoner of Azkaban', 0.7309308611870757],
 ['Harry Potter and the Order of the Phoenix', 0.7217362181392408],
 ['Harry Potter and the Deathly Hallows', 0.7181677376484684],
 ['The Magical Worlds of Harry Potter', 0.7146800943719254],
 ['Harry Potter and the Cursed Child', 0.6725872668915877],
 ['The Ickabog', 0.6218310147923186],
 ['Fantastic Beasts and Where to Find Them', 0.6161251907593163]]

 The Hobbit recommendations:
[['The History of The Hobbit', 0.78046806361336],
 ['The Lord of the Rings', 0.764041360399863],
 ['The Annotated Hobbit', 0.7444487700381719],
 ['The Marvellous Land of Snergs', 0.6904192459951058],
 ['The Children of Húrin', 0.6804096398917605],
 ['The Road to Middle-Earth', 0.6596135627601877],
 ['Mr. Bliss', 0.6543540064849226],
 ['The Silmarillion', 0.640755416461898],
 ['J. R. R. Tolkien: A Biography', 0.6391232063030203],
 ['Tolkien: Maker of Middle-earth', 0.6309609890944725]]

 Harry Potter and the Philosopher's Stone and The Hobbit recommendations:
[['Harry Potter and the Half-Blood Prince', 0.6018217616032179],
 ['Harry Potter and the Prisoner of Azkaban', 0.5989788027468591],
 ['The Magical Worlds of Harry Potter', 0.5909785871728664],
 ['Harry Potter and the Order of the Phoenix', 0.5889168038270771],
 ['The Lord of the Rings', 0.5881581367207107],
 ['Harry Potter and the Chamber of Secrets', 0.5868542056295735],
 ['Harry Potter and the Deathly Hallows', 0.5805140956814785],
 ['The Weirdstone of Brisingamen', 0.5725139741586933],
 ['The Children of Húrin', 0.5661655486061915],
 ['Harry Potter and the Goblet of Fire', 0.5653645423523244]]

To-Do ↩

• Devising methods to best combine recommendations for more than one input (including an option for distinterest)
• Adding arguments to data_utils.input_conversion_dict() based on Wikipedia languages to simplify parsing arguments
• Adding and refining methods for recommendations in wikirec.model
• Creating, improving and sharing examples
• Compiling other sources of open-source data that can be used to augment input data
  • Potentially writing scripts to load this data for significant topics
• Updates to wikirec.languages as lemmatization and other linguistic package dependencies evolve
• Allowing euclidean distance measurements for LDA based recommendations in wikirec.model.gen_sim_matrix()
• Expanding the documentation
• Improving tests for greater code coverage

References

List of references

• https://towardsdatascience.com/building-a-recommendation-system-using-neural-network-embeddings-1ef92e5c80c9

• https://towardsdatascience.com/wikipedia-data-science-working-with-the-worlds-largest-encyclopedia-c08efbac5f5c

Powered By