Can my computer make jokes

Anyone interested in understanding how Word2Vec Recurrent Neural Networks work? What about doing it while teaching your computer to make jokes?

Diogo Pinto, a Data Scientist at Feedzai is showing you how to do this. Diogo calls himself a Machine Learning, Security and Scala Enthusiast, multipotentialite, parkour practitioner and a full-time wisdom seeker.

The talks is aimed at anyone with an introductory level of knowledge of Deep Learning, that want to know more about how Recurrent Neural Nets works.

P.S.: Be warned, computer jokes will be bad, but imagine it like your weird friend that fills in the silence with the driest jokes ever.

Transcript

1. Can my computer make jokes? by Diogo Pinto 2018-11-28
2. Can my computer make jokes? by Diogo Pinto 2018-11-28
3. About me • Trained as Software Engineer • Grown into a Data Scientist generalist • Other interests • Blockchain • Distributed systems • (Cyber- and non-cyber-) Security • Psychology and Philosophy • Kung Fu and Parkour practitioner #nofilters @diogojapinto /in/diogojapinto/ diogojapinto@gmail.com
4. @diogojapinto /in/diogojapinto/ diogojapinto@gmail.com About me • Trained as Software Engineer • Grown into a Data Scientist generalist • Other interests • Blockchain • Distributed systems • (Cyber- and non-cyber-) Security • Psychology and Philosophy • Kung Fu and Parkour practitioner #nofilters
5. @diogojapinto /in/diogojapinto/ diogojapinto@gmail.com About me • Trained as Software Engineer • Grown into a Data Scientist generalist • Other interests • Blockchain • Distributed systems • (Cyber- and non-cyber-) Security • Psychology and Philosophy • Kung Fu and Parkour practitioner #nofilters
6. The plan for today What is the role of Humor? How can a computer use words? Does my computer has better sense of humor than I do? What are the take-aways?
7. “Know your audience Luke”
8. “Know your audience Luke” • Who here is aware of recent Machine Learning achievements?
9. “Know your audience Luke” • Who here is aware of recent Machine Learning achievements? • Who here has some intuition behind the inner workings of Neural Networks (aka Differentiable Programming)?
10. “Know your audience Luke” • Who here is aware of recent Machine Learning achievements? • Who here has some intuition behind the inner workings of Neural Networks (aka Differentiable Programming)? • Who here has a dark sense of humor?
11. What is the role of Humor? “A day without laughter is a day wasted” Charlie Chaplin
12. History of Humor [1] [1] The First Joke: Exploring the Evolutionary Origins of Humor
13. History of Humor [1] • Humor is complex and a high cognitive function [1] The First Joke: Exploring the Evolutionary Origins of Humor
14. History of Humor [1] • Humor is complex and a high cognitive function • Nuanced verbal phrasing + Prevailing social dynamics [1] The First Joke: Exploring the Evolutionary Origins of Humor
15. History of Humor [1] • Humor is complex and a high cognitive function • Nuanced verbal phrasing + Prevailing social dynamics • Can combine language skills, theory-of-mind, symbolism, abstract thinking, social perception… [1] The First Joke: Exploring the Evolutionary Origins of Humor
16. History of Humor [1] • Humor is complex and a high cognitive function • Nuanced verbal phrasing + Prevailing social dynamics • Can combine language skills, theory-of-mind, symbolism, abstract thinking, social perception… • The basic ability seems “instinctive” [1] The First Joke: Exploring the Evolutionary Origins of Humor
17. History of Humor [1] • Humor is complex and a high cognitive function • Nuanced verbal phrasing + Prevailing social dynamics • Can combine language skills, theory-of-mind, symbolism, abstract thinking, social perception… • The basic ability seems “instinctive” • Ubiquitous and Universal [1] The First Joke: Exploring the Evolutionary Origins of Humor
18. History of Humor [1] • Humor is complex and a high cognitive function • Nuanced verbal phrasing + Prevailing social dynamics • Can combine language skills, theory-of-mind, symbolism, abstract thinking, social perception… • The basic ability seems “instinctive” • Ubiquitous and Universal • It is probably coded somehow in our genetic code • People laugh without appreciation for the causal factors [1] The First Joke: Exploring the Evolutionary Origins of Humor
19. History of Humor [1] • Humor is complex and a high cognitive function • Nuanced verbal phrasing + Prevailing social dynamics • Can combine language skills, theory-of-mind, symbolism, abstract thinking, social perception… • The basic ability seems “instinctive” • Ubiquitous and Universal • It is probably coded somehow in our genetic code • People laugh without appreciation for the causal factors • Humor dates back thousands of years [1] The First Joke: Exploring the Evolutionary Origins of Humor
20. History of Humor [1] • Humor is complex and a high cognitive function • Nuanced verbal phrasing + Prevailing social dynamics • Can combine language skills, theory-of-mind, symbolism, abstract thinking, social perception… • The basic ability seems “instinctive” • Ubiquitous and Universal • It is probably coded somehow in our genetic code • People laugh without appreciation for the causal factors • Humor dates back thousands of years • Greek “laughing philosopher” Democritus [1] The First Joke: Exploring the Evolutionary Origins of Humor
21. History of Humor [1] • Humor is complex and a high cognitive function • Nuanced verbal phrasing + Prevailing social dynamics • Can combine language skills, theory-of-mind, symbolism, abstract thinking, social perception… • The basic ability seems “instinctive” • Ubiquitous and Universal • It is probably coded somehow in our genetic code • People laugh without appreciation for the causal factors • Humor dates back thousands of years • Greek “laughing philosopher” Democritus • Humor conversations observed in Australian aboriginals • Lived genetically isolated for at least 35000 years [1] The First Joke: Exploring the Evolutionary Origins of Humor
22. Why Humor is relevant for Machine Learning[1] • Humor is complex and a high cognitive function • Nuanced verbal phrasing + Prevailing social dynamics • Can combine language skills, theory-of-mind, symbolism, abstract thinking, social perception… • The basic ability seems “instinctive” • Ubiquitous and Universal • It is probably coded somehow in our genetic code • People laugh without appreciation for the causal factors • Humor dates back thousands of years • Greek “laughing philosopher” Democritus • Humor conversations observed in Australian aboriginals • Lived genetically isolated for at least 35000 years [1] The First Joke: Exploring the Evolutionary Origins of Humor
23. Let’s look at the data
24. Let’s look at the data • Dataset of short jokes from Kaggle user avmoudgil95
25. Let’s look at the data • Dataset of short jokes from Kaggle user avmoudgil95 • E.g.: • It's crazy how my ex was so upset about losing me that he had to build a life with a new woman.
26. Let’s look at the data • Dataset of short jokes from Kaggle user avmoudgil95 • E.g.: • It's crazy how my ex was so upset about losing me that he had to build a life with a new woman.
27. Let’s look at the data • Dataset of short jokes from Kaggle user avmoudgil95 • E.g.: • It's crazy how my ex was so upset about losing me that he had to build a life with a new woman. • Where does Noah keep his bees? In the Ark Hives
28. Let’s look at the data • Dataset of short jokes from Kaggle user avmoudgil95 • E.g.: • It's crazy how my ex was so upset about losing me that he had to build a life with a new woman. • Where does Noah keep his bees? In the Ark Hives
29. Let’s look at the data • Dataset of short jokes from Kaggle user avmoudgil95 • E.g.: • It's crazy how my ex was so upset about losing me that he had to build a life with a new woman. • Where does Noah keep his bees? In the Ark Hives • What sex position produces the ugliest children? Ask your mother.
30. Let’s look at the data • Dataset of short jokes from Kaggle user avmoudgil95 • E.g.: • It's crazy how my ex was so upset about losing me that he had to build a life with a new woman. • Where does Noah keep his bees? In the Ark Hives • What sex position produces the ugliest children? Ask your mother.
31. Let’s look at the data • Dataset of short jokes from Kaggle user avmoudgil95 • E.g.: • It's crazy how my ex was so upset about losing me that he had to build a life with a new woman. • Where does Noah keep his bees? In the Ark Hives • What sex position produces the ugliest children? Ask your mother. • Chuck Norris doesn't have blood. He is filled with magma.
32. Let’s look at the data • Dataset of short jokes from Kaggle user avmoudgil95 • E.g.: • It's crazy how my ex was so upset about losing me that he had to build a life with a new woman. • Where does Noah keep his bees? In the Ark Hives • What sex position produces the ugliest children? Ask your mother. • Chuck Norris doesn't have blood. He is filled with magma.
33. Let’s look at the data • Dataset of short jokes from Kaggle user avmoudgil95 • E.g.: • It's crazy how my ex was so upset about losing me that he had to build a life with a new woman. • Where does Noah keep his bees? In the Ark Hives • What sex position produces the ugliest children? Ask your mother. • Chuck Norris doesn't have blood. He is filled with magma. • Length:
34. How can a computer use words? “A picture may be worth a thousand words, but well chosen, words will take you where pictures never can” Unknown
35. Why does representation matter?
36. Why does representation matter? Sender country risky “fee” count Spam Detector Machine Learning Model Words count
37. Why does representation matter? Sender country risky “fee” count Spam Detector Machine Learning Model Words count Sender country risky “fee” count Words count Spam? Email 1 True 0 203 True Email 2 False 2 345 False Email 3 True 10 180 True
38. Why does representation matter? Sender country risky “fee” count Spam Detector Machine Learning Model Words count Sender country risky “fee” count Words count “prince” count Spam? Email 1 True 0 203 5 True Email 2 False 2 345 0 False Email 3 True 10 180 3 True
39. Why does representation matter? Sender country risky “fee” count Spam Detector Machine Learning Model Words count Sender country risky “fee” count Words count Spam? Email 1 True 0 203 True Email 2 False 2 345 False Email 3 True 10 180 True
40. Why does representation matter? Sender country risky “fee” count Spam Detector Machine Learning Model Words count Sender country risky “fee” count Spam? Email 1 True 0 True Email 2 False 2 False Email 3 True 10 True
41. Why does representation matter? Sender country risky “fee” count Spam Detector Machine Learning Model Words count Sender country risky “fee” count Words count Spam? Email 1 True 0 203 True Email 2 False 2 345 False Email 3 True 10 180 True Data entries should be comparable and consistent
42. The problem of Words Representation
43. The problem of Words Representation • An array of characters, each one a byte • Example: “I like beer a lot” “I own a lot of wine” • Representation: 49206c696b6520626565722061206c6f74 49206f776e2061206c6f74206f662077696e65
44. The problem of Words Representation • An array of characters, each one a byte • Variable length • Difficult comparison between entries • Example: “I like beer a lot” “I own a lot of wine” • Representation: 49206c696b6520626565722061206c6f74 49206f776e2061206c6f74206f662077696e65
45. The problem of Words Representation • An array of characters, each one a byte • Variable length • Difficult comparison between entries • Bag-of-Words • Example: “I like beer a lot” “I own a lot of wine” • Representation: I like beer a lot own of wine 1 1 1 1 1 0 0 0 1 0 0 1 1 1 1 1
46. The problem of Words Representation • An array of characters, each one a byte • Variable length • Difficult comparison between entries • Bag-of-Words • By discarding order we are able to generalize • Example: “I like beer a lot” “I own a lot of wine” • Representation: I like beer a lot own of wine 1 1 1 1 1 0 0 0 1 0 0 1 1 1 1 1
47. The problem of Words Representation • An array of characters, each one a byte • Variable length • Difficult comparison between entries • Bag-of-Words • By discarding order we are able to generalize • Term Frequency – Inverse Document Frequency • Example: “I like beer a lot” “I own a lot of wine” • Representation: I like beer a lot own of wine 0.1 0.3 0.6 0.1 0.4 0 0 0 0.1 0 0 0.1 0.4 0.5 0.1 0.7
48. The problem of Words Representation • An array of characters, each one a byte • Variable length • Difficult comparison between entries • Bag-of-Words • By discarding order we are able to generalize • Term Frequency – Inverse Document Frequency • Words frequency in document • Example: “I like beer a lot” “I own a lot of wine” • Representation: I like beer a lot own of wine 0.1 0.3 0.6 0.1 0.4 0 0 0 0.1 0 0 0.1 0.4 0.5 0.1 0.7
49. The problem of Words Representation • An array of characters, each one a byte • Variable length • Difficult comparison between entries • Bag-of-Words • By discarding order we are able to generalize • Term Frequency – Inverse Document Frequency • Words frequency in document • Rarity of words across documents • Example: “I like beer a lot” “I own a lot of wine” • Representation: I like beer a lot own of wine 0.1 0.3 0.6 0.1 0.4 0 0 0 0.1 0 0 0.1 0.4 0.5 0.1 0.7
50. The problem of Words Representation • An array of characters, each one a byte • Variable length • Difficult comparison between entries • Bag-of-Words • By discarding order we are able to generalize • Term Frequency – Inverse Document Frequency • Words frequency in document • Rarity of words across documents • Example: “I like beer a lot” “I own a lot of wine” • Representation: I like beer a lot own of wine 0.1 0.3 0.6 0.1 0.4 0 0 0 0.1 0 0 0.1 0.4 0.5 0.1 0.7
51. The problem of Words Representation • An array of characters, each one a byte • Variable length • Difficult comparison between entries • Bag-of-Words • By discarding order we are able to generalize • Term Frequency – Inverse Document Frequency • Words frequency in document • Rarity of words across documents • Example: “I like beer a lot” “I own a lot of wine” • Representation: I like beer a lot own of wine 0.1 0.3 0.6 0.1 0.4 0 0 0 0.1 0 0 0.1 0.4 0.5 0.1 0.7
52. What about semantics?
53. What about semantics? Doc I like beer a lot own of wine beer wine own
54. What about semantics? • Meaning of words is lost Doc I like beer a lot own of wine beer 0 0 1 0 0 0 0 0 wine 0 0 0 0 0 0 0 1 own 0 0 0 0 0 1 0 0
55. What about semantics? • Meaning of words is lost • Distance(wine, beer) = Distance(wine, own) wine own beer Doc I like beer a lot own of wine beer 0 0 1 0 0 0 0 0 wine 0 0 0 0 0 0 0 1 own 0 0 0 0 0 1 0 0
56. What about semantics? • Meaning of words is lost • Distance(wine, beer) = Distance(wine, own) • Distributed representations can help
57. What about semantics? • Meaning of words is lost • Distance(wine, beer) = Distance(wine, own) • Distributed representations can help Doc D1 D2 beer 0.2 0.7 wine 0.1 0.8 own 0.8 0.1
58. What about semantics? • Meaning of words is lost • Distance(wine, beer) = Distance(wine, own) • Distributed representations can help Doc D1 D2 beer 0.2 0.7 wine 0.1 0.8 own 0.8 0.1 0 0,3 0,5 0,8 1 0 0,2 0,4 0,6 0,8 wine own beer
59. What about semantics? • Meaning of words is lost • Distance(wine, beer) = Distance(wine, own) • Distributed representations can help • Reduce the dimensionality footprint • Semantics encoded as “proximity” Doc D1 D2 beer 0.2 0.7 wine 0.1 0.8 own 0.8 0.1 0 0,3 0,5 0,8 1 0 0,2 0,4 0,6 0,8 wine own beer
60. What about semantics? • Meaning of words is lost • Distance(wine, beer) = Distance(wine, own) • Distributed representations can help • Reduce the dimensionality footprint • Semantics encoded as “proximity” • Word2Vec • Start with “random” word representations with dimension d
61. What about semantics? • Meaning of words is lost • Distance(wine, beer) = Distance(wine, own) • Distributed representations can help • Reduce the dimensionality footprint • Semantics encoded as “proximity” • Word2Vec • Start with “random” word representations with dimension d I [0.8, 0.12, …] like … beer … a … lot …
62. What about semantics? • Meaning of words is lost • Distance(wine, beer) = Distance(wine, own) • Distributed representations can help • Reduce the dimensionality footprint • Semantics encoded as “proximity” • Word2Vec • Start with “random” word representations with dimension d • From the representation of a given word predict a randomly sampled context word
63. What about semantics? • Meaning of words is lost • Distance(wine, beer) = Distance(wine, own) • Distributed representations can help • Reduce the dimensionality footprint • Semantics encoded as “proximity” • Word2Vec • Start with “random” word representations with dimension d • From the representation of a given word predict a randomly sampled context word I like beer a lot
64. What about semantics? • Meaning of words is lost • Distance(wine, beer) = Distance(wine, own) • Distributed representations can help • Reduce the dimensionality footprint • Semantics encoded as “proximity” • Word2Vec • Start with “random” word representations with dimension d • From the representation of a given word predict a randomly sampled context word I like beer a lot • Training examples: • (I, like)
65. What about semantics? • Meaning of words is lost • Distance(wine, beer) = Distance(wine, own) • Distributed representations can help • Reduce the dimensionality footprint • Semantics encoded as “proximity” • Word2Vec • Start with “random” word representations with dimension d • From the representation of a given word predict a randomly sampled context word I like beer a lot • Training examples: • (I, like) • (like, I) • (like, beer)
66. What about semantics? • Meaning of words is lost • Distance(wine, beer) = Distance(wine, own) • Distributed representations can help • Reduce the dimensionality footprint • Semantics encoded as “proximity” • Word2Vec • Start with “random” word representations with dimension d • From the representation of a given word predict a randomly sampled context word I like beer a lot • Training examples: • (I, like) • (like, I) • (like, beer) • (beer, like) • (beer, a)
67. What about semantics? • Meaning of words is lost • Distance(wine, beer) = Distance(wine, own) • Distributed representations can help • Reduce the dimensionality footprint • Semantics encoded as “proximity” • Word2Vec • Start with “random” word representations with dimension d • From the representation of a given word predict a randomly sampled context word I like beer a lot • Training examples: • (I, like) • (like, I) • (like, beer) • (beer, like) • (beer, a) • (a, beer) • (a, lot)
68. What about semantics? • Meaning of words is lost • Distance(wine, beer) = Distance(wine, own) • Distributed representations can help • Reduce the dimensionality footprint • Semantics encoded as “proximity” • Word2Vec • Start with “random” word representations with dimension d • From the representation of a given word predict a randomly sampled context word I like beer a lot • Training examples: • (I, like) • (like, I) • (like, beer) • (beer, like) • (beer, a) • (a, beer) • (a, lot) • (lot, a)
69. What about semantics? • Meaning of words is lost • Distance(wine, beer) = Distance(wine, own) • Distributed representations can help • Reduce the dimensionality footprint • Semantics encoded as “proximity” • Word2Vec • Start with “random” word representations with dimension d • From the representation of a given word predict a randomly sampled context word I [0.8, 0.12, …] like … beer … a … lot …
70. What about semantics? • Meaning of words is lost • Distance(wine, beer) = Distance(wine, own) • Distributed representations can help • Reduce the dimensionality footprint • Semantics encoded as “proximity” • Word2Vec • Start with “random” word representations with dimension d • From the representation of a given word predict a randomly sampled context word I [0.8, 0.12, …] like … beer … a … lot …
71. What about semantics? • Meaning of words is lost • Distance(wine, beer) = Distance(wine, own) • Distributed representations can help • Reduce the dimensionality footprint • Semantics encoded as “proximity” • Word2Vec • Start with “random” word representations with dimension d • From the representation of a given word predict a randomly sampled context word I [0.8, 0.12, …] like … beer … a … lot … Shady Mathy Stuff
72. What about semantics? • Meaning of words is lost • Distance(wine, beer) = Distance(wine, own) • Distributed representations can help • Reduce the dimensionality footprint • Semantics encoded as “proximity” • Word2Vec • Start with “random” word representations with dimension d • From the representation of a given word predict a randomly sampled context word I [0.8, 0.12, …] like … beer … a … lot … Shady Mathy Stuff Score(I) Score(like) Score(beer) … …
73. What about semantics? • Meaning of words is lost • Distance(wine, beer) = Distance(wine, own) • Distributed representations can help • Reduce the dimensionality footprint • Semantics encoded as “proximity” • Word2Vec • Start with “random” word representations with dimension d • From the representation of a given word predict a randomly sampled context word • Change the word representation in the direction that helps predicting the context word I [0.8, 0.12, …] like … beer … a … lot … Shady Mathy Stuff Score(I) Score(like) Score(beer) … …
74. What about semantics? • Meaning of words is lost • Distance(wine, beer) = Distance(wine, own) • Distributed representations can help • Reduce the dimensionality footprint • Semantics encoded as “proximity” • Word2Vec • Start with “random” word representations with dimension d • From the representation of a given word predict a randomly sampled context word • Change the word representation in the direction that helps predicting the context word I [0.7, 0.15, …] like … beer … a … lot … Shady Mathy Stuff Error(I) Error(like) Error(beer) … …
75. But I’m here to see Python code
76. But I’m here to see Python code >>> jokes[20] 
 "Why do you never see elephants hiding in trees? 'Cause they are freaking good at it"
77. But I’m here to see Python code • Pre-Processing • Transform into word lists >>> import nltk >>> nltk.download('punkt') >>> tokenizer = nltk.data.load('english.pickle') >>> jokes_sentences = tokenizer.tokenize_sents(jokes) >>> jokes_sentences[20] 
 ['Why do you never see elephants hiding in trees?', "'Cause they are freaking good at it"]
78. But I’m here to see Python code >>> def sentence_to_wordlist(raw): ... clean = re.sub(r'[^a-zA-Z]', ' ', raw) ... words = clean.split() ... words_lower = [w.lower() for w in words] ... return words_lower 
 >>> jokes_word_lists = [ [sentence_to_wordlist(str(s)) for s in joke if len(s) > 0] for joke in jokes_sentences] 
 >>> jokes_word_lists[20] 
 [['why', 'do', 'you', 'never', 'see', 'elephants', 'hiding', 'in', 'trees'], ['cause', 'they', 'are', 'freaking', 'good', 'at', 'it']] • Pre-Processing • Transform into word lists
79. But I’m here to see Python code >>> from nltk.corpus import stopwords >>> nltk.download('stopwords') >>> stop_ws = set(stopwords.words('english')) >>> jokes_non_stop_word_lists = [[[word for word in s if word not in stop_ws] for s in joke] for joke in jokes_word_lists] >>> jokes_non_stop_word_lists[20] 
 [['never', 'see', 'elephants', 'hiding', 'trees'], ['cause', 'freaking', 'good']] • Pre-Processing • Transform into word lists • Remove Stop Words
80. But I’m here to see Python code >>> from nltk.stem import PorterStemmer >>> porter = PorterStemmer() >>> jokes_stemmed = [[[porter.stem(w) for w in s] for s in joke] for joke in jokes_non_stop_word_lists] >>> jokes_stemmed[20] 
 [['never', 'see', 'eleph', 'hide', 'tree'], ['caus', 'freak', 'good']] • Pre-Processing • Transform into word lists • Remove Stop Words • Stemming
81. But I’m here to see Python code >>> jokes_flatten = [s for j in jokes_stemmed for s in j] >>> num_features = 300 # dimensionality of the resulting word vectors >>> min_word_count = 3 # minimum word count threshold >>> num_workers = multiprocessing.cpu_count() # number of threads to # run in parallel >>> context_size = 7 # context window length >>> downsampling = 1e-3 # downlsample setting for frequent words >>> seed = 42 # seed for the rng, to make the results reproducible >>> sg = 1 # skip-gram (instead of CBOW) 
 >>> jokes2vec = w2v.Word2Vec( ... sg=sg, ... seed=seed, ... workers=num_workers, ... size=num_features, ... min_count=min_word_count, ... window=context_size, ... sample=downsampling ... ) >>> jokes2vec.build_vocab(jokes_flatten) >>> jokes2vec.train(jokes_flatten, total_examples=jokes2vec.corpus_count, epochs=jokes2vec.iter) • Pre-Processing • Transform into word lists • Remove Stop Words • Stemming • Obtain word embeddings
82. But I’m here to see Python code >>> import seaborn as sns >>> sns.set(style="darkgrid") >>> from sklearn.manifold import TSNE >>> tsne = TSNE(n_components=2, random_state=seed) >>> all_word_vectors_matrix = jokes2vec.wv.syn0 >>> all_word_vectors_matrix_2d = tsne.fit_transform(all_word_vectors_matrix) >>> points = pd.DataFrame( ... [ ... (word, coords[0], coords[1]) ... for word, coords in [ ... (word, all_word_vectors_matrix_2d[jokes2vec.wv.vocab[word].index]) ... for word in jokes2vec.wv.vocab ... ] ... ], ... columns=['word', 'x', 'y'] ... ) 
 >>> points.plot.scatter('x', 'y', s=10, figsize=(20, 12), alpha=0.6) • Pre-Processing • Transform into word lists • Remove Stop Words • Stemming • Obtain word embeddings • Visualize
83. But I’m here to see Python code • Pre-Processing • Transform into word lists • Remove Stop Words • Stemming • Obtain word embeddings • Visualize
84. But I’m here to see Python code • Pre-Processing • Transform into word lists • Remove Stop Words • Stemming • Obtain word embeddings • Visualize
85. But I’m here to see Python code • Pre-Processing • Transform into word lists • Remove Stop Words • Stemming • Obtain word embeddings • Visualize
86. But I’m here to see Python code • Pre-Processing • Transform into word lists • Remove Stop Words • Stemming • Obtain word embeddings • Visualize
87. But I’m here to see Python code • Pre-Processing • Transform into word lists • Remove Stop Words • Stemming • Obtain word embeddings • Visualize
88. But I’m here to see Python code • Pre-Processing • Transform into word lists • Remove Stop Words • Stemming • Obtain word embeddings • Visualize
89. But I’m here to see Python code • Pre-Processing • Transform into word lists • Remove Stop Words • Stemming • Obtain word embeddings • Visualize
90. But I’m here to see Python code • Pre-Processing • Transform into word lists • Remove Stop Words • Stemming • Obtain word embeddings • Visualize • Embeddings “algebra”
91. But I’m here to see Python code >>> jokes2vec.most_similar('facebook’) [('fb', 0.7791515588760376), ('unfriend', 0.7512669563293457), ('status', 0.7433165907859802), ('myspac', 0.7160271406173706), ('notif', 0.6782281398773193), ('retweet', 0.6745551824569702), ('timelin', 0.672653079032898), ('twitter', 0.6709973812103271), ('privaci', 0.6695473194122314), ('linkedin', 0.6655823588371277)] • Pre-Processing • Transform into word lists • Remove Stop Words • Stemming • Obtain word embeddings • Visualize • Embeddings “algebra” • Semantic similarity
92. But I’m here to see Python code >>> jokes2vec.most_similar('lol’) [('lmao', 0.7619101405143738), ('tho', 0.7015952467918396), ('haha', 0.6999001502990723), ('hahaha', 0.6714984178543091), ('omg', 0.6711198091506958), ('pl', 0.6587743163108826), ('bc', 0.6558701992034912), ('gona', 0.6529208421707153), ('ppl', 0.6476595401763916), ('yea', 0.6466178894042969)] • Pre-Processing • Transform into word lists • Remove Stop Words • Stemming • Obtain word embeddings • Visualize • Embeddings “algebra” • Semantic similarity
93. But I’m here to see Python code >>> jokes2vec.most_similar('sex’) [('anal', 0.5785183906555176), ('unprotect', 0.5359092950820923), ('foreplay', 0.5343884825706482), ('brussel', 0.5324864387512207), ('foursom', 0.5289731025695801), ('twosom', 0.5187283158302307), ('threesom', 0.5119856595993042), ('geneticist', 0.5064876079559326), ('intercours', 0.5030955076217651), ('oral', 0.5015446543693542)] • Pre-Processing • Transform into word lists • Remove Stop Words • Stemming • Obtain word embeddings • Visualize • Embeddings “algebra” • Semantic similarity
94. But I’m here to see Python code >>> def nearest_similarity_cosmul(start1, end1, end2): ... similarities = jokes2vec.most_similar_cosmul( ... positive=[end2, start1], ... negative=[end1] ... ) ... start2 = similarities[0][0] ... print('{start1} is related to {end1}, as {start2} is related to {end2}'.format(**locals())) >>> nearest_similarity_cosmul('dude', 'man', 'woman’) "dude is related to man, as chick is related to woman" • Pre-Processing • Transform into word lists • Remove Stop Words • Stemming • Obtain word embeddings • Visualize • Embeddings “algebra” • Semantic similarity • Linear relationships
95. Does my computer has better sense of humor than I do? “Comparing yourself to others is an act of violence against the self” Ivanla Vanzant
96. Traditional Neural Networks
97. Traditional Neural Networks • Input and output lengths are mostly pre-determined and static
98. Traditional Neural Networks ML Model Input Output • Input and output lengths are mostly pre-determined and static
99. Traditional Neural Networks • Input and output lengths are mostly pre-determined and static • Weak for modelling sequences ML Model Input Output
100. Traditional Neural Networks • Input and output lengths are mostly pre-determined and static • Weak for modelling sequences • Recurrent Neural Networks
101. Traditional Neural Networks • Input and output lengths are mostly pre-determined and static • Weak for modelling sequences • Recurrent Neural Networks • Enable the model to keep a memory between executions
102. Traditional Neural Networks • Input and output lengths are mostly pre-determined and static • Weak for modelling sequences • Recurrent Neural Networks • Enable the model to keep a memory between executions ML Model Input Output
103. Traditional Neural Networks • Input and output lengths are mostly pre-determined and static • Weak for modelling sequences • Recurrent Neural Networks • Enable the model to keep a memory between executions ML Model Input Output
104. Traditional Neural Networks • Input and output lengths are mostly pre-determined and static • Weak for modelling sequences • Recurrent Neural Networks • Enable the model to keep a memory between executions
105. Traditional Neural Networks • Input and output lengths are mostly pre-determined and static • Weak for modelling sequences • Recurrent Neural Networks • Enable the model to keep a memory between executions ML Model Input 1 Output 1
106. Traditional Neural Networks • Input and output lengths are mostly pre-determined and static • Weak for modelling sequences • Recurrent Neural Networks • Enable the model to keep a memory between executions ML Model Input 1 Output 1
107. Traditional Neural Networks • Input and output lengths are mostly pre-determined and static • Weak for modelling sequences • Recurrent Neural Networks • Enable the model to keep a memory between executions ML Model Input 1 Output 1 ML Model Input 2 Output 2
108. Traditional Neural Networks • Input and output lengths are mostly pre-determined and static • Weak for modelling sequences • Recurrent Neural Networks • Enable the model to keep a memory between executions ML Model Input 1 Output 1 ML Model Input 2 Output 2 ML Model Input 3 Output 3
109. Traditional Neural Networks • Input and output lengths are mostly pre-determined and static • Weak for modelling sequences • Recurrent Neural Networks • Enable the model to keep a memory between executions ML Model Ø I ML Model I like ML Model like beer
110. Traditional Neural Networks • Input and output lengths are mostly pre-determined and static • Weak for modelling sequences • Recurrent Neural Networks • Enable the model to keep a memory between executions ML Model ML Model ML Model • RNNs are very versatile
111. Traditional Neural Networks • Input and output lengths are mostly pre-determined and static • Weak for modelling sequences • Recurrent Neural Networks • Enable the model to keep a memory between executions ML Model Input 1 Output 1 ML Model Input 2 Output 2 ML Model Input 3 Output 3 • RNNs are very versatile • Many to many
112. Traditional Neural Networks • Input and output lengths are mostly pre-determined and static • Weak for modelling sequences • Recurrent Neural Networks • Enable the model to keep a memory between executions ML Model Input 1 ML Model Input 2 ML Model Input 3 Output 1 • RNNs are very versatile • Many to many • Many to one
113. Traditional Neural Networks • Input and output lengths are mostly pre-determined and static • Weak for modelling sequences • Recurrent Neural Networks • Enable the model to keep a memory between executions ML Model Input 1 Output 1 ML Model Output 2 ML Model Output 3 • RNNs are very versatile • Many to many • Many to one • One to many
114. AI-Powered Jokes
115. AI-Powered Jokes • How? Generate character by character
116. AI-Powered Jokes ML Model b ML Model e • How? Generate character by character
117. AI-Powered Jokes ML Model b ML Model e e • How? Generate character by character
118. AI-Powered Jokes ML Model b ML Model e ML Model e • How? Generate character by character
119. AI-Powered Jokes • How? Generate character by character ML Model b ML Model e ML Model e r
120. AI-Powered Jokes • How? Generate character by character • A seed is provides context for the network ML Model b ML Model e ML Model e r
121. Stop poking my brain and show 
 me code!!! • Pre-Processing
122. Stop poking my brain and show 
 me code!!! • Pre-Processing • Character 㲗 Integer Index >>> jokes_concat = ''.join(jokes) >>> chars = sorted(list(set(jokes_concat))) >>> idx2char = chars >>> char2idx = {c: i for i, c in enumerate(idx2char)}
123. Stop poking my brain and show 
 me code!!! • Pre-Processing • Character 㲗 Integer Index • Training set preparation >>> jokes_sample = random.sample(jokes, math.floor(len(jokes) * 0.5)) >>> maxlen = 40 >>> step = 3 >>> sentences = [] >>> next_chars = [] >>> for joke in jokes_sample: >>> if len(joke) < (maxlen + 1): >>> continue >>> for i in range(0, len(joke) - maxlen, step): >>> sentences.append(joke[i:i+maxlen]) >>> next_chars.append(joke[i+maxlen])
124. Stop poking my brain and show 
 me code!!! • Pre-Processing • Character 㲗 Integer Index • Training set preparation • Convert into integers >>> x = np.zeros((len(sentences), maxlen, len(chars)), dtype=np.bool) >>> y = np.zeros((len(sentences), len(chars)), dtype=np.bool) >>> for i, sentence in enumerate(sentences): >>> for j, char in enumerate(sentence): >>> x[i, j, char2idx[char]] = 1. >>> y[i, char2idx[next_chars[i]]] = 1. >>> idx = np.random.permutation(len(x)) >>> x = x[idx] >>> y = y[idx]
125. Stop poking my brain and show 
 me code!!! • Pre-Processing • Character 㲗 Integer Index • Training set preparation • Convert into integers • Define the model >>> from keras.models import Sequential >>> from keras.layers import Dense, Dropout, LSTM >>> model = Sequential() >>> model.add(LSTM(256, ... input_shape=(maxlen, len(chars)), ... return_sequences=True)) >>> model.add(Dropout(0.2)) >>> model.add(LSTM(256)) >>> model.add(Dropout(0.2)) >>> model.add(Dense(len(chars), activation='softmax')) >>> model.compile(loss='categorical_crossentropy’, ... optimizer='adam’, ... metrics=['categorical_crossentropy'])
126. Stop poking my brain and show 
 me code!!! • Pre-Processing • Character 㲗 Integer Index • Training set preparation • Convert into integers • Define the model • Train the model >>> model.fit(x, y, batch_size=128, epochs=8)
127. Stop poking my brain and show 
 me code!!! • Pre-Processing • Character 㲗 Integer Index • Training set preparation • Convert into integers • Define the model • Train the model >>> model.fit(x, y, batch_size=128, epochs=8)
128. Are the jokes any good? • Using the start of an existing joke as seed
129. Are the jokes any good? • Using the start of an existing joke as seed Example: How many dead hookers does it take to sc
130. Are the jokes any good? • Using the start of an existing joke as seed Example: How many dead hookers does it take to screw in a light bulb? Two, but it was the same time.
131. Are the jokes any good? • Using the start of an existing joke as seed • Sentence structure is surprisingly good Example: How many dead hookers does it take to screw in a light bulb? Two, but it was the same time.
132. Are the jokes any good? • Using the start of an existing joke as seed • Sentence structure is surprisingly good • Coherence with the answer is a bit lacking Example: How many dead hookers does it take to screw in a light bulb? Two, but it was the same time.
133. Are the jokes any good? • Using the start of an existing joke as seed • Sentence structure is surprisingly good • Coherence with the answer is a bit lacking • A weirder example
134. Are the jokes any good? • Using the start of an existing joke as seed • Sentence structure is surprisingly good • Coherence with the answer is a bit lacking • A weirder example Example: Am I the only one who closes the silverw
135. Are the jokes any good? • Using the start of an existing joke as seed • Sentence structure is surprisingly good • Coherence with the answer is a bit lacking • A weirder example Example: Am I the only one who closes the silverwine costume have sex with Dram the ground? Prostitute too scoring out wrenking out running into a teeth in other people? Are you surprised whereed? Chocolate Good!!!!!!!!!!!!? Don't you buy me? nit-te
136. Are the jokes any good? • Using the start of an existing joke as seed • Sentence structure is surprisingly good • Coherence with the answer is a bit lacking • A weirder example • Custom seed
137. Are the jokes any good? • Using the start of an existing joke as seed • Sentence structure is surprisingly good • Coherence with the answer is a bit lacking • A weirder example • Custom start • A letter Example: D
138. Are the jokes any good? • Using the start of an existing joke as seed • Sentence structure is surprisingly good • Coherence with the answer is a bit lacking • A weirder example • Custom start • A letter Example: Do you have a beer from starbucks? An asshole in the back of the back, but I leave the same time. I think I was all day. The barman says "I don't know what the fucking card move in the back of his life
139. Are the jokes any good? • Using the start of an existing joke as seed • Sentence structure is surprisingly good • Coherence with the answer is a bit lacking • A weirder example • Custom start • A letter • “What do you call” jokes
140. Are the jokes any good? • Using the start of an existing joke as seed • Sentence structure is surprisingly good • Coherence with the answer is a bit lacking • A weirder example • Custom start • A letter • “What do you call” jokes Examples: • What do you call a nut on house? A coint in his circus. • What do you call a bill on the oven? A condom. • What do you call a blowjob? A social storm. • What do you call a shit of country in a car? A lifetime experience. • What do you call a dog device? A garbanzo bean with a curry. • What do you call a disappointment on the bathroom? A pilot, you can't take a shit. • What do you call a dialogast? A farmer. • What do you call a dog first? A sandwich. • What do you call a dick on his shoes? A woman in the stairs.
141. What are the take- aways? “Art is never finished, only abandoned” Leonardo Da Vinci
142. Today I learned Topics covered •Distributed representations are great for categorical fields •Word2Vec and LSTM rules •Python rules Current trends •Divide and Conquer •Attention •Generative Adversarial Networks •Sequence-to-Sequence •RNN-based Variational Auto Encoder Potential paths of exploration •Try using attention and longer spans of memory. •Active learning classifier for jokes quality •Do the same for motivational quotes
143. And what now? I’m a Software Engineer Fast.ai Machine Learning for Coders Seek problems on your surrounding and do a POC Python rules I am a Data Scientist DeepLearning.ai Deep Learning Specialization Develop Probability and Statistics •Udacity Intro to Descriptive Statistics •Udacity Intro to Inferential Statistics I went through the wrong door and noticed free beer Enjoy it :D
144. Bibliography • https://en.wikipedia.org/wiki/Theories_of_humor • https://en.wikipedia.org/wiki/Humor_research • https://en.wikipedia.org/wiki/Computational_humor • https://www.iflscience.com/technology/ais-attempts- at-oneliner-jokes-are-unintentionally-hilarious/ • https://motherboard.vice.com/en_us/article/z43nke/ joke-telling-robots-are-the-final-frontier-of-artificial- intelligence • https://medium.com/@davidolarinoye/will-ai-ever-be- able-to-make-a-joke-808a656b53a6 • https://journals.sagepub.com/doi/pdf/ 10.1177/147470490600400129 • https://towardsdatascience.com/word2vec-skip-gram- model-part-1-intuition-78614e4d6e0b • https://www.analyticsvidhya.com/blog/2017/06/word- embeddings-count-word2veec/ • https://www.kaggle.com/abhinavmoudgil95/short- jokes • https://arxiv.org/pdf/1708.02709.pdf • https://journals.sagepub.com/doi/pdf/ 10.1177/147470490600400129 • http://blog.aylien.com/overview-word-embeddings- history-word2vec-cbow-glove/
145. Thank you @diogojapinto /in/diogojapinto/ diogojapinto@gmail.com