Word Embeddings for Natural Language Processing in Python @ PyMunich meetup

Joint "Hacking Machine Learning" (@hack_ai) and PyMunich (@PyMunich) meetup in Munich, Germany

Transcript

1. Word Embeddings 
 for NLP in Python Marco Bonzanini
 PyMunich // Hacking Machine Learning 20th June 2017
2. Nice to meet you
3. WORD EMBEDDINGS?
4. Word Embeddings Word Vectors Distributed Representations = =
5. Why should you care?
6. Why should you care? Data representation
 is crucial
7. Applications
8. Applications • Classification / tagging • Recommendation Systems • Search Engines (Query Expansion) • Machine Translation
9. One-hot Encoding
10. One-hot Encoding Rome Paris Italy France = [1, 0, 0, 0, 0, 0, …, 0] = [0, 1, 0, 0, 0, 0, …, 0] = [0, 0, 1, 0, 0, 0, …, 0] = [0, 0, 0, 1, 0, 0, …, 0]
11. One-hot Encoding Rome Paris Italy France = [1, 0, 0, 0, 0, 0, …, 0] = [0, 1, 0, 0, 0, 0, …, 0] = [0, 0, 1, 0, 0, 0, …, 0] = [0, 0, 0, 1, 0, 0, …, 0] Rome Paris word V
12. One-hot Encoding Rome Paris Italy France = [1, 0, 0, 0, 0, 0, …, 0] = [0, 1, 0, 0, 0, 0, …, 0] = [0, 0, 1, 0, 0, 0, …, 0] = [0, 0, 0, 1, 0, 0, …, 0] V = vocabulary size (huge)
13. Bag-of-words
14. Bag-of-words doc_1 doc_2 … doc_N = [32, 14, 1, 0, …, 6] = [ 2, 12, 0, 28, …, 12] … = [13, 0, 6, 2, …, 0]
15. Bag-of-words doc_1 doc_2 … doc_N = [32, 14, 1, 0, …, 6] = [ 2, 12, 0, 28, …, 12] … = [13, 0, 6, 2, …, 0] Rome Paris word V
16. Word Embeddings
17. Word Embeddings Rome Paris Italy France = [0.91, 0.83, 0.17, …, 0.41] = [0.92, 0.82, 0.17, …, 0.98] = [0.32, 0.77, 0.67, …, 0.42] = [0.33, 0.78, 0.66, …, 0.97]
18. Word Embeddings Rome Paris Italy France = [0.91, 0.83, 0.17, …, 0.41] = [0.92, 0.82, 0.17, …, 0.98] = [0.32, 0.77, 0.67, …, 0.42] = [0.33, 0.78, 0.66, …, 0.97] n. dimensions << vocabulary size
19. Word Embeddings Rome Paris Italy France = [0.91, 0.83, 0.17, …, 0.41] = [0.92, 0.82, 0.17, …, 0.98] = [0.32, 0.77, 0.67, …, 0.42] = [0.33, 0.78, 0.66, …, 0.97]
20. Word Embeddings Rome Paris Italy France = [0.91, 0.83, 0.17, …, 0.41] = [0.92, 0.82, 0.17, …, 0.98] = [0.32, 0.77, 0.67, …, 0.42] = [0.33, 0.78, 0.66, …, 0.97]
21. Word Embeddings Rome Paris Italy France = [0.91, 0.83, 0.17, …, 0.41] = [0.92, 0.82, 0.17, …, 0.98] = [0.32, 0.77, 0.67, …, 0.42] = [0.33, 0.78, 0.66, …, 0.97]
22. Word Embeddings Rome Paris Italy France
23. Word Embeddings is-capital-of
24. Word Embeddings Paris + Italy - France ≈ Rome Rome
25. THE MAIN INTUITION
26. Distributional Hypothesis
27. –J.R. Firth 1957 “You shall know a word 
 by the company it keeps.”
28. –Z. Harris 1954 “Words that occur in similar context tend to have similar meaning.”
29. Context ≈ Meaning
30. I enjoyed eating some pizza at the restaurant
31. I enjoyed eating some pizza at the restaurant Word
32. I enjoyed eating some pizza at the restaurant The company it keeps Word
33. I enjoyed eating some pizza at the restaurant I enjoyed eating some weißwurst at the restaurant
34. I enjoyed eating some pizza at the restaurant I enjoyed eating some weißwurst at the restaurant
35. I enjoyed eating some pizza at the restaurant I enjoyed eating some weißwurst at the restaurant Same context
36. I enjoyed eating some pizza at the restaurant I enjoyed eating some weißwurst at the restaurant Pizza = Weisswurst ?! Same context
37. A BIT OF THEORY word2vec
40. Vector Calculation
41. Vector Calculation Goal: learn vec(word)
42. Vector Calculation Goal: learn vec(word) 1. Choose objective function
43. Vector Calculation Goal: learn vec(word) 1. Choose objective function 2. Init: random vectors
44. Vector Calculation Goal: learn vec(word) 1. Choose objective function 2. Init: random vectors 3. Run gradient descent
45. I enjoyed eating some pizza at the restaurant
46. I enjoyed eating some pizza at the restaurant
47. I enjoyed eating some pizza at the restaurant
48. I enjoyed eating some pizza at the restaurant Maximise the likelihood 
 of the context given the focus word
49. I enjoyed eating some pizza at the restaurant Maximise the likelihood 
 of the context given the focus word P(i | pizza) P(enjoyed | pizza) … P(restaurant | pizza)
50. Example I enjoyed eating some pizza at the restaurant
51. I enjoyed eating some pizza at the restaurant Iterate over context words Example
52. I enjoyed eating some pizza at the restaurant bump P( i | pizza ) Example
53. I enjoyed eating some pizza at the restaurant bump P( enjoyed | pizza ) Example
54. I enjoyed eating some pizza at the restaurant bump P( eating | pizza ) Example
55. I enjoyed eating some pizza at the restaurant bump P( some | pizza ) Example
56. I enjoyed eating some pizza at the restaurant bump P( at | pizza ) Example
57. I enjoyed eating some pizza at the restaurant bump P( the | pizza ) Example
58. I enjoyed eating some pizza at the restaurant bump P( restaurant | pizza ) Example
59. I enjoyed eating some pizza at the restaurant Move to next focus word and repeat Example
60. I enjoyed eating some pizza at the restaurant bump P( i | at ) Example
61. I enjoyed eating some pizza at the restaurant bump P( enjoyed | at ) Example
62. I enjoyed eating some pizza at the restaurant … you get the picture Example
63. P( eating | pizza )
64. P( eating | pizza ) ??
65. P( eating | pizza ) Input word Output word
66. P( eating | pizza ) Input word Output word P( vec(eating) | vec(pizza) )
67. P( vout | vin ) P( vec(eating) | vec(pizza) ) P( eating | pizza ) Input word Output word
68. P( vout | vin ) P( vec(eating) | vec(pizza) ) P( eating | pizza ) Input word Output word ???
69. P( vout | vin )
70. cosine( vout, vin )
71. cosine( vout, vin ) [-1, 1]
72. softmax(cosine( vout, vin ))
73. softmax(cosine( vout, vin )) [0, 1]
74. softmax(cosine( vout, vin )) P (vout | vin) = exp(cosine(vout , vin)) P k 2 V exp(cosine(vk , vin))
75. Vector Calculation Recap
76. Vector Calculation Recap Learn vec(word)
77. Vector Calculation Recap Learn vec(word) by gradient descent
78. Vector Calculation Recap Learn vec(word) by gradient descent on the softmax probability
79. Plot Twist
82. Paragraph Vector a.k.a. doc2vec i.e. P(vout | vin, label)
83. A BIT OF PRACTICE
85. pip install gensim
86. Case Study 1: Skills and CVs
87. Case Study 1: Skills and CVs Data set of ~300k resumes Each experience is a “sentence” Each experience has 3-15 skills Approx 15k unique skills
88. Case Study 1: Skills and CVs from gensim.models import Word2Vec fname = 'candidates.jsonl' corpus = ResumesCorpus(fname) model = Word2Vec(corpus)
89. Case Study 1: Skills and CVs model.most_similar('chef') [('cook', 0.94), ('bartender', 0.91), ('waitress', 0.89), ('restaurant', 0.76), ...]
90. Case Study 1: Skills and CVs model.most_similar('chef',
 negative=['food']) [('puppet', 0.93), ('devops', 0.92), ('ansible', 0.79), ('salt', 0.77), ...]
91. Case Study 1: Skills and CVs Useful for: Data exploration Query expansion/suggestion Recommendations
92. Case Study 2: Beer!
93. Case Study 2: Beer! Data set of ~2.9M beer reviews 89 different beer styles 635k unique tokens 185M total tokens
94. Case Study 2: Beer! from gensim.models import Doc2Vec fname = 'ratebeer_data.csv' corpus = RateBeerCorpus(fname) model = Doc2Vec(corpus)
95. Case Study 2: Beer! from gensim.models import Doc2Vec fname = 'ratebeer_data.csv' corpus = RateBeerCorpus(fname) model = Doc2Vec(corpus) 3.5h on my laptop … remember to pickle
96. Case Study 2: Beer! model.docvecs.most_similar('Stout') [('Sweet Stout', 0.9877), ('Porter', 0.9620), ('Foreign Stout', 0.9595), ('Dry Stout', 0.9561), ('Imperial/Strong Porter', 0.9028), ...]
97. Case Study 2: Beer! model.most_similar([model.docvecs['Stout']]) 
 [('coffee', 0.6342), ('espresso', 0.5931), ('charcoal', 0.5904), ('char', 0.5631), ('bean', 0.5624), ...]
98. Case Study 2: Beer! model.most_similar([model.docvecs['Wheat Ale']]) 
 [('lemon', 0.6103), ('lemony', 0.5909), ('wheaty', 0.5873), ('germ', 0.5684), ('lemongrass', 0.5653), ('wheat', 0.5649), ('lime', 0.55636), ('verbena', 0.5491), ('coriander', 0.5341), ('zesty', 0.5182)]
99. PCA
100. Dark beers
101. Strong beers
102. Sour beers
103. Lagers
104. Wheat beers
105. Case Study 2: Beer! Useful for: Understanding the language of beer enthusiasts Planning your next pint Classification
106. FINAL REMARKS
107. But we’ve been
 doing this for X years
108. But we’ve been
 doing this for X years • Approaches based on co-occurrences are not new • Think SVD / LSA / LDA • … but they are usually outperformed by word2vec • … and don’t scale as well as word2vec
109. Efficiency
110. Efficiency • There is no co-occurrence matrix
 (vectors are learned directly) • Softmax has complexity O(V)
 Hierarchical Softmax only O(log(V))
111. Garbage in, garbage out
112. Garbage in, garbage out • Pre-trained vectors are useful • … until they’re not • The business domain is important • The pre-processing steps are important • > 100K words? Maybe train your own model • > 1M words? Yep, train your own model
113. Summary
114. Summary • Word Embeddings are magic! • Big victory of unsupervised learning • Gensim makes your life easy
115. Credits & Readings
116. Credits & Readings Credits • Lev Konstantinovskiy (@gensim_py) • Chris E. Moody (@chrisemoody) see videos on lda2vec Readings • Deep Learning for NLP (R. Socher) http://cs224d.stanford.edu/ • “word2vec parameter learning explained” by Xin Rong More readings • “GloVe: global vectors for word representation” by Pennington et al. • “Dependency based word embeddings” and “Neural word embeddings as implicit matrix factorization” by O. Levy and Y. Goldberg
117. THANK YOU @MarcoBonzanini GitHub.com/bonzanini marcobonzanini.com