Upgrade to Pro — share decks privately, control downloads, hide ads and more …

Document Classification with word2vec at PyData...

Document Classification with word2vec at PyData NYC

Lev Konstantinovskiy

May 25, 2016
Tweet

More Decks by Lev Konstantinovskiy

Other Decks in Technology

Transcript

  1. Word Embeddings for Fun and Profit Lev Konstantinovskiy Community Manager

    at Gensim @teagermylk http://rare-technologies.com/
  2. About Lev Konstantinovskiy @teagermylk https://github.com/tmylk Graduate school drop-out in Algebraic

    Geometry Worked in Trading IT Graduate of Galvanize Data Science Bootcamp in San Francisco Community manager and consultant at RaRe Technologies. NLP consulting and open-source development of Natural Language Processing packages in Python.
  3. Part 1. Word2vec theory. Part 2. Document classification. Practical stuff.

    Notebook at: http://small.cat/dry (also in meetup page comments)
  4. The business problem to be solved in Part 2 You

    run a movie studio. Every day you receive thousands of proposals for movies to make. Need to send them to the right department for consideration! One department per genre. Need to classify plots by genre.
  5. Word embeddings can be used for: - recommendation engines -

    automated text tagging (this talk) - synonyms and search query expansion - machine translation - plain feature engineering
  6. What is a word embedding? ‘Word embedding’ = ‘word vectors’

    = ‘distributed representations’ It is a dense representation of words in a low-dimensional vector space. One-hot representation: king = [1 0 0 0.. 0 0 0 0 0] queen = [0 1 0 0 0 0 0 0 0] book = [0 0 1 0 0 0 0 0 0] king = [0.9457, 0.5774, 0.2224] Distributed representation, Word-embedding, Word-vector:
  7. word2vec relies on the “Distributional hypothesis”: “You shall know a

    word by the company it keeps” -J. R. Firth 1957 Richard Socher’s NLP course http://cs224d.stanford.edu/lectures/CS224d-Lecture2.pdf How to come up with an embedding?
  8. For the theory, take Richard Sochers’s class Richard Socher’s NLP

    course http://cs224d.stanford.edu/lectures/CS224d-Lecture2.pdf
  9. “The fox jumped over the lazy dog” Maximize the likelihood

    of seeing the context words given the word over. P(the|over) P(fox|over) P(jumped|over) P(the|over) P(lazy|over) P(dog|over) word2vec algorithm Used with permission from @chrisemoody http://www.slideshare.net/ChristopherMoody3/word2vec-lda-and- introducing-a-new-hybrid-algorithm-lda2vec
  10. Probability should depend on the word vectors. Used with permission

    from @chrisemoody http://www.slideshare.net/ChristopherMoody3/word2vec-lda-and- introducing-a-new-hybrid-algorithm-lda2vec P(fox|over) P(v fox |v over )
  11. A twist: two vectors for every word Used with permission

    from @chrisemoody http://www.slideshare.net/ChristopherMoody3/word2vec-lda-and- introducing-a-new-hybrid-algorithm-lda2vec Should depend on whether it’s the input or the output. P(v OUT |v IN ) “The fox jumped over the lazy dog” v IN
  12. Twist: two vectors for every word Used with permission from

    @chrisemoody http://www.slideshare.net/ChristopherMoody3/word2vec-lda-and- introducing-a-new-hybrid-algorithm-lda2vec Should depend on whether it’s the input or the output. P(v OUT |v IN ) “The fox jumped over the lazy dog” v IN v OUT = P(v THE |v OVER )
  13. Twist: two vectors for every word Used with permission from

    @chrisemoody http://www.slideshare.net/ChristopherMoody3/word2vec-lda-and- introducing-a-new-hybrid-algorithm-lda2vec Should depend on whether it’s the input or the output. P(v OUT |v IN ) “The fox jumped over the lazy dog” v IN v OUT
  14. Twist: two vectors for every word Used with permission from

    @chrisemoody http://www.slideshare.net/ChristopherMoody3/word2vec-lda-and- introducing-a-new-hybrid-algorithm-lda2vec Should depend on whether it’s the input or the output. P(v OUT |v IN ) “The fox jumped over the lazy dog” v IN v OUT
  15. Twist: two vectors for every word Used with permission from

    @chrisemoody http://www.slideshare.net/ChristopherMoody3/word2vec-lda-and- introducing-a-new-hybrid-algorithm-lda2vec Should depend on whether it’s the input or the output. P(v OUT |v IN ) “The fox jumped over the lazy dog” v IN v OUT
  16. Twist: two vectors for every word Used with permission from

    @chrisemoody http://www.slideshare.net/ChristopherMoody3/word2vec-lda-and- introducing-a-new-hybrid-algorithm-lda2vec Should depend on whether it’s the input or the output. P(v OUT |v IN ) “The fox jumped over the lazy dog” v IN v OUT
  17. Twist: two vectors for every word Used with permission from

    @chrisemoody http://www.slideshare.net/ChristopherMoody3/word2vec-lda-and- introducing-a-new-hybrid-algorithm-lda2vec Should depend on whether it’s the input or the output. P(v OUT |v IN ) “The fox jumped over the lazy dog” v IN v OUT
  18. Twist: two vectors for every word Used with permission from

    @chrisemoody http://www.slideshare.net/ChristopherMoody3/word2vec-lda-and- introducing-a-new-hybrid-algorithm-lda2vec Should depend on whether it’s the input or the output. P(v OUT |v IN ) “The fox jumped over the lazy dog” v IN v OUT
  19. Twist: two vectors for every word Used with permission from

    @chrisemoody http://www.slideshare.net/ChristopherMoody3/word2vec-lda-and- introducing-a-new-hybrid-algorithm-lda2vec Should depend on whether it’s the input or the output. P(v OUT |v IN ) “The fox jumped over the lazy dog” v IN v OUT
  20. How to define P(v OUT |v IN )? First, define

    similarity. How similar are two vectors? Just dot product for unit length vectors v OUT * v IN Used with permission from @chrisemoody http://www.slideshare.net/ChristopherMoody3/word2vec-lda-and- introducing-a-new-hybrid-algorithm-lda2vec
  21. Get a probability in [0,1] out of similarity in [-1,

    1] Normalization term over all out words Used with permission from @chrisemoody http://www.slideshare.net/ChristopherMoody3/word2vec-lda-and- introducing-a-new-hybrid-algorithm-lda2vec
  22. Word2vec is big victory of unsupervised learning Google ran word2vec

    on 100billion of unlabelled documents. Then shared their trained model. Thanks to Google for cutting our training time to zero!. :)
  23. What is the genre of this plot? In a future

    world devastated by disease, a convict is sent back in time to gather information about the man-made virus that wiped out most of the human population on the planet.
  24. Of course it is SCI-FI In a future world devastated

    by disease, a convict is sent back in time to gather information about the man-made virus that wiped out most of the human population on the planet.
  25. What is the genre of this one? Mrs. Dashwood and

    her three daughters are left in straitened circumstances. When Elinor forms an attachment for the wealthy Edward Ferrars, his family disapproves and separates them. And though Mrs. Jennings tries to match the worthy (and rich) Colonel Brandon to her, Marianne finds the dashing and fiery John Willoughby more to her taste.
  26. ROMANCE When Mr. Dashwood dies, he must leave the bulk

    of his estate to the son by his first marriage, which leaves his second wife and their three daughters (Elinor, Marianne, and Margaret) in straitened circumstances. They are taken in by a kindly cousin, but their lack of fortune affects the marriageability of both practical Elinor and romantic Marianne. When Elinor forms an attachment for the wealthy Edward Ferrars, his family disapproves and separates them. And though Mrs. Jennings tries to match the worthy (and rich) Colonel Brandon to her, Marianne finds the dashing and fiery John Willoughby more to her taste. Both relationships are sorely tried.
  27. When Mr. Dashwood dies, he must leave the bulk of

    his estate to the son by his first marriage, which leaves his second wife and their three daughters (Elinor, Marianne, and Margaret) in straitened circumstances. They are taken in by a kindly cousin... romance In a future world devastated by disease, a convict is sent back in time to gather information about the man- made virus that wiped out most of the human population on the planet. sci-fi The text is very different so should be some signal there
  28. “Hello World in 7 different models” - Will show APIs

    to run these models without tuning: - Bag of words - Character n-grams - TF-IDF - Averaging word2vec vectors - doc2vec - Deep IR - Word Mover's Distance - No tuning today. To show how to tune them will require 7 more talks!
  29. Simple baseline: 47% TF-IDF just counting words in a document,

    adjusting for doc length, word frequency and word-doc frequency Counting words, char n-grams are similar: 42, 44. Embedding Classifier Accuracy Train time,s Predict time, s Baseline tf-idf bag of words Logistic 47 2 1 TfidfVectorizer( min_df=2, tokenizer=nltk.word_tokenize, preprocessor=None,stop_words='english') Disclaimer: First run, “out of the box” performance. No tuning.
  30. Word2vec to... A Document Classifier We need some features to

    power our favourite classifier (logistic regression/KNN) We have vectors for words but need vectors for documents. How to create a document classifier out of a set of word vectors? For KNN, how similar is one sequence of words to another sequence of words?
  31. Averaging word vectors aka ‘Naive document vector’ Just add word

    vectors together! All words in a book ‘A tale of two cities’ Should add up to ‘class-struggle’ Mike Tamir https://www.linkedin.com/pulse/short-introduction-using-word2vec-text-classification-mike
  32. Averaging word vectors: 52 % accuracy # loading 1.5 GB

    archive into memory wv = Word2Vec.load_word2vec_format( "GoogleNews-vectors-negative300.bin.gz") X_train_naive_dv = naived2v_conversion_np(wv,flat_list_train) logreg = linear_model.LogisticRegression(n_jobs=-1, C=1e5) logreg = logreg.fit(X_train_naive_dv, train_data['tag']) Embedding Classifier Accuracy Train time,s Predict time 250 docs, s Averaging word vectors from pre- trained Google News Logistic 52 2 (thanks to Google!) 1 Disclaimer: First run, “out of the box” performance. No tuning.
  33. Introducing Doc2vec Used with permission from @chrisemoody http://www.slideshare.net/ChristopherMoody3/word2vec-lda-and- introducing-a-new-hybrid-algorithm-lda2vec Tag

    is ‘a word that is in every context in the doc’ P(v OUT |v IN ,COMEDY) “The fox jumped over the lazy dog. (COMEDY)” v IN v OUT = P(v FOX |v OVER ,v COMEDY )
  34. Introducing Doc2vec Used with permission from @chrisemoody http://www.slideshare.net/ChristopherMoody3/word2vec-lda-and- introducing-a-new-hybrid-algorithm-lda2vec Tag

    is ‘a word that is in every context in the doc’ P(v OUT |v IN ,COMEDY) “The fox jumped over the lazy dog. (COMEDY)” v IN v OUT = P(v JUMPED |v OVER ,v COMEDY )
  35. Doc2vec DM Tag is ‘a word that is in every

    context in the doc’ Le, Mikolov. Distributed Representations of Sentences and Documents 2014
  36. Closest words to sci-fi model.most_similar([mdm_alt.docvecs['sci-fi']]) [[('alien', 0.4514704942703247), ('express', 0.4008052945137024), ('space',

    0.40043187141418457), ('planet', 0.3805035352706909), ('ant', 0.37011784315109253), ('ferocious', 0.36217403411865234), ('ship', 0.35579410195350647), ('hole', 0.3422626256942749) ]
  37. Closest words romance model.most_similar([mdm_alt.docvecs['romance']]) [('say', 0.38082122802734375), ('skill', 0.3159002363681793), ('leads', 0.3063559830188751),

    ('local', 0.3018215596675873), ('millionaire', 0.2863730788230896), ('located', 0.28458985686302185), ('hood', 0.2830425798892975), ('heir', 0.2802196145057678), ('died', 0.27215155959129333), ('indians', 0.26776593923568726)]
  38. Doc2vec: 52 % # simple gensim doc2vec api model =

    Doc2Vec( trainsent, workers=2, size=100, iter=20, dm=1) train_targets, train_regressors = zip(*[ (doc.tags[0], model.infer_vector(doc.words, steps=20)) for doc in trainsent]) Embedding Classifier Accuracy Train time,s Predict time, s Doc2vec Logistic 52 25 1 Disclaimer: First run, “out of the box” performance. No tuning.
  39. Embedding Classifier Accuracy Train time,s Predict time on 250 docs,

    s Baseline tf-idf bag of words Logistic 47 2 1 Averaging word vectors from pre- trained Google News Logistic 52 2 (thanks to Google!) 1 Doc2vec Logistic 52 25 1 Bayesian Inversion Max likelihood 30 120 1 Word Movers Distance on Google News KNN 42 1 (thanks to Google!) 1800 Comparing “Hello-worlds”. Rough, first run, “out of the box” performance. No tuning.
  40. No neural network magic out of the box:( Simple baselines

    are not much worse than fancy methods. For magic and tuned models - see many many academic papers
  41. Which model is easiest to tune and debug? “What caused

    this error?” “What is wrong with comedy genre?” “What can I do to fix this class of errors?”
  42. Embedding Classifier Accuracy Train time,s Predict time, s Debug/tune Baseline

    tf-idf bag of words Logistic 47 2 1 Easy Averaging word vectors from pre- trained Google News Logistic 52 2 (thanks to Google!) 1 Hard Doc2vec Logistic 52 25 1 Hard Bayesian Inversion Max likelihood 30 120 1 Hard Word Movers Distance on Google News KNN 42 1 (thanks to Google!) 1800 Hard Comparing “Hello-worlds”. Rough, first run, “out of the box” performance. No tuning.
  43. There is a place for word embeddings - Small improvement

    matters (maybe you want your paper accepted) - Lots of data - Extra features to existing ensembles - Need directionality (analogies)
  44. RARE Training •customized, interactive corporate training hosted on-site for technical

    teams of 5-15 developers, engineers, analysts and data scientists •2-day intensives include Python Best Practices and Practical Machine Learning, and 1-day intensive Topic Modelling RNDr. Radim Řehůřek, Ph.D. Gordon Mohr, BA in CS & Econ industry-leading instructors for more information email training@rare-technologies. com
  45. Thanks! Lev Konstantinovskiy github.com/tmylk @teagermylk Come and learn about word2vec

    with us! Events coming up: - Tutorials sprint at PyCon in Portland, Oregon 2- 6 June - New York Corporate Training 13 - 17 June