Deep Learning for Natural Language Processing: Word Embeddings

1. 1 @graphiﬁc Roelof Pieters Deep Learning for Natural Language Processing: Word Embeddings 3 December 2015   KTH www.csc.kth.se/~roelof/ roelof@kth.se

2. Language Understanding 2

3. Can we understand Language ? 1. Language is ambiguous:  Every sentence has many possible interpretations. 2. Language is productive:  We will always encounter new words or new constructions 3. Language is culturally speciﬁc Some of the challenges in Language Understanding: 3

4. Can we understand Language ? 1. Language is ambiguous:  Every sentence has many possible interpretations. 2. Language is productive:  We will always encounter new words or new constructions • plays well with others VB ADV P NN NN NN P DT • fruit ﬂies like a banana NN NN VB DT NN NN VB P DT NN NN NN P DT NN NN VB VB DT NN • the students went to class DT NN VB P NN 4 Some of the challenges in Language Understanding:

5. Can we understand Language ? 1. Language is ambiguous:  Every sentence has many possible interpretations. 2. Language is productive:  We will always encounter new words or new constructions 5 Some of the challenges in Language Understanding:

6. [Karlgren 2014, NLP Sthlm Meetup]6

7. Can we understand Language ? 1. Language is ambiguous:  Every sentence has many possible interpretations. 2. Language is productive:  We will always encounter new words or new constructions 3. Language is culturally speciﬁc Some of the challenges in Language Understanding: 7

8. ML: Traditional Approach 1. Gather as much LABELED data as you can get 2. Throw some algorithms at it (mainly put in an SVM and keep it at that) 3. If you actually have tried more algos: Pick the best 4. Spend hours hand engineering some features / feature selection / dimensionality reduction (PCA, SVD, etc) 5. Repeat… For each new problem/question:: 8

9. Machine Learning for NLP Data Classic Approach: Data is fed into a learning algorithm: Learning   Algorithm 9

10. Machine Learning for NLP some of the (many) treebank datasets source: http://www-nlp.stanford.edu/links/statnlp.html#Treebanks ! 10

11. Penn Treebank That’s a lot of “manual” work: 11

12. • the students went to class DT NN VB P NN • plays well with others VB ADV P NN NN NN P DT • fruit ﬂies like a banana NN NN VB DT NN NN VB P DT NN NN NN P DT NN NN VB VB DT NN With a lot of issues: Penn Treebank 12

13. Machine Learning for NLP Learning   Algorithm Data “Features” Prediction Prediction/  Classiﬁer train set test set 13

14. Machine Learning for NLP Learning   Algorithm “Features” Prediction Prediction/  Classiﬁer train set test set 14

15. One Model rules them all ?    DL approaches have been successfully applied to: Deep Learning: Why for NLP ? Automatic summarization Coreference resolution Discourse analysis Machine translation Morphological segmentation Named entity recognition (NER) Natural language generation Natural language understanding Optical character recognition (OCR) Part-of-speech tagging Parsing Question answering Relationship extraction sentence boundary disambiguation Sentiment analysis Speech recognition Speech segmentation Topic segmentation and recognition Word segmentation Word sense disambiguation Information retrieval (IR) Information extraction (IE) Speech processing 15

16. Deep Learning: Why for NLP ? 16

17. • What is the meaning of a word?  (Lexical semantics) • What is the meaning of a sentence?  ([Compositional] semantics) • What is the meaning of a longer piece of text? (Discourse semantics) Semantics: Meaning 18

18. • NLP treats words mainly (rule-based/statistical approaches at least) as atomic symbols:  • or in vector space:  • also known as “one hot” representation. • Its problem ? Word Representation Love Candy Store [0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 …] Candy [0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 …] AND Store [0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 …] = 0 ! 19

19. Word Representation 20

20. • Structure corresponds to meaning: Structure and Meaning 21

21. • Semantics • Syntax 22 NLP: what can we work with?

22. • Language models deﬁne probability distributions over (natural language) strings or sentences • Joint and Conditional Probability Language Model 23

23. • Language models deﬁne probability distributions over (natural language) strings or sentences Language Model 24

24. • Language models deﬁne probability distributions over (natural language) strings or sentences Language Model 25

25. Word senses What is the meaning of words? • Most words have many different senses:  dog = animal or sausage? How are the meanings of different words related? • - Specific relations between senses:  Animal is more general than dog. • - Semantic fields:  money is related to bank 26

26. Word senses Polysemy: • A lexeme is polysemous if it has different related senses • bank = financial institution or building Homonyms: • Two lexemes are homonyms if their senses are unrelated, but they happen to have the same spelling and pronunciation • bank = (financial) bank or (river) bank 27

27. Word senses: relations Symmetric relations: • Synonyms: couch/sofa  Two lemmas with the same sense • Antonyms: cold/hot, rise/fall, in/out  Two lemmas with the opposite sense Hierarchical relations: • Hypernyms and hyponyms: pet/dog  The hyponym (dog) is more speciﬁc than the hypernym (pet) • Holonyms and meronyms: car/wheel  The meronym (wheel) is a part of the holonym (car) 28

28. Distributional representations “You shall know a word by the company it keeps”  (J. R. Firth 1957) One of the most successful ideas of modern statistical NLP! these words represent banking • Hard (class based) clustering models • Soft clustering models 29

29. Distributional hypothesis He ﬁlled the wampimuk, passed it around and we all drunk some We found a little, hairy wampimuk sleeping behind the tree (McDonald & Ramscar 2001) 30

30. Distributional semantics Landauer and Dumais (1997), Turney and Pantel (2010), … 31

31. Distributional semantics Distributional meaning as co-occurrence vector: 32

32. Distributional representations • Taking it further: • Continuous word embeddings • Combine vector space semantics with the prediction of probabilistic models • Words are represented as a dense vector: Candy = 33

33. Word Embeddings: SocherVector Space Model adapted rom Bengio, “Representation Learning and Deep Learning”, July, 2012, UCLA In a perfect world: 34

34. Word Embeddings: SocherVector Space Model adapted rom Bengio, “Representation Learning and Deep Learning”, July, 2012, UCLA In a perfect world: the country of my birth the place where I was born 35

35. • Can theoretically (given enough units) approximate “any” function • and fit to “any” kind of data • Efficient for NLP: hidden layers can be used as word lookup tables • Dense distributed word vectors + efficient NN training algorithms: • Can scale to billions of words ! Why Neural Networks for NLP? 36

36. • Representation of words as continuous vectors has a long history (Hinton et al. 1986; Rumelhart et al. 1986; Elman 1990) • First neural network language model: NNLM (Bengio et al. 2001; Bengio et al. 2003) based on earlier ideas of distributed representations for symbols (Hinton 1986) How? 37

37. Word Embeddings: SocherVector Space Model Figure (edited) from Bengio, “Representation Learning and Deep Learning”, July, 2012, UCLA In a perfect world: the country of my birth the place where I was born ? … 38

38. Compositionality Principle of compositionality: the “meaning (vector) of a complex expression (sentence) is determined by: — Gottlob Frege   (1848 - 1925) - the meanings of its constituent expressions (words) and - the rules (grammar) used to combine them” 39

39. • How do we handle the compositionality of language in our models? 40 Compositionality

40. • How do we handle the compositionality of language in our models? • Recursion :  the same operator (same parameters) is applied repeatedly on diﬀerent components 41 Compositionality

41. • How do we handle the compositionality of language in our models? • Option 1: Recurrent Neural Networks (RNN) 42 RNN 1: Recurrent Neural Networks

42. • How do we handle the compositionality of language in our models? • Option 2: Recursive Neural Networks (also sometimes called RNN) 43 RNN 2: Recursive Neural Networks

43. • achieved SOTA in 2011 on Language Modeling (WSJ AR task) (Mikolov et al., INTERSPEECH 2011): • and again at ASRU 2011: 44 Recurrent Neural Networks “Comparison to other LMs shows that RNN LMs are state of the art by a large margin. Improvements inrease with more training data.” “[ RNN LM trained on a] single core on 400M words in a few days, with 1% absolute improvement in WER on state of the art setup” Mikolov, T., Karaﬁat, M., Burget, L., Cernock, J.H., Khudanpur, S. (2011)  Recurrent neural network based language model

44. 45 Recurrent Neural Networks (simple recurrent   neural network for LM) input hidden layer(s) output layer + sigmoid activation function + softmax function: Mikolov, T., Karaﬁat, M., Burget, L., Cernock, J.H., Khudanpur, S. (2011)  Recurrent neural network based language model

45. 46 Recurrent Neural Networks backpropagation through time

46. 47 Recurrent Neural Networks backpropagation through time class based recurrent NN [code (Mikolov’s RNNLM Toolkit) and more info: http://rnnlm.org/ ]

47. • Recursive Neural Network for LM (Socher et al. 2011; Socher 2014) • achieved SOTA on new Stanford Sentiment Treebank dataset (but comparing it to many other models): Recursive Neural Network 48 Socher, R., Perelygin,, A., Wu, J.Y., Chuang, J., Manning, C.D., Ng, A.Y., Potts, C. (2013)  Recursive Deep Models for Semantic Compositionality Over a Sentiment Treebank info & code: http://nlp.stanford.edu/sentiment/

48. Recursive Neural Tensor Network 49 code & info: http://www.socher.org/index.php/Main/ ParsingNaturalScenesAndNaturalLanguageWithRecursiveNeuralNetworks Socher, R., Liu, C.C., NG, A.Y., Manning, C.D. (2011)   Parsing Natural Scenes and Natural Language with Recursive Neural Networks

49. Recursive Neural Tensor Network 50

50. • RNN (Socher et al. 2011a) Recursive Neural Network 51 Socher, R., Perelygin,, A., Wu, J.Y., Chuang, J., Manning, C.D., Ng, A.Y., Potts, C. (2013)  Recursive Deep Models for Semantic Compositionality Over a Sentiment Treebank info & code: http://nlp.stanford.edu/sentiment/

51. • RNN (Socher et al. 2011a) • Matrix-Vector RNN (MV-RNN) (Socher et al., 2012) Recursive Neural Network 52 Socher, R., Perelygin,, A., Wu, J.Y., Chuang, J., Manning, C.D., Ng, A.Y., Potts, C. (2013)  Recursive Deep Models for Semantic Compositionality Over a Sentiment Treebank info & code: http://nlp.stanford.edu/sentiment/

52. • RNN (Socher et al. 2011a) • Matrix-Vector RNN (MV-RNN) (Socher et al., 2012) • Recursive Neural Tensor Network (RNTN) (Socher et al. 2013) Recursive Neural Network 53

53. • negation detection: Recursive Neural Network 54 Socher, R., Perelygin,, A., Wu, J.Y., Chuang, J., Manning, C.D., Ng, A.Y., Potts, C. (2013)  Recursive Deep Models for Semantic Compositionality Over a Sentiment Treebank info & code: http://nlp.stanford.edu/sentiment/

54. NP PP/IN NP DT NN PRP$ NN Parse Tree Recurrent NN for Vector Space 55

55. NP PP/IN NP DT NN PRP$ NN Parse Tree INDT NN PRP NN Compositionality 56 Recurrent NN: CompositionalityRecurrent NN for Vector Space

56. NP IN NP PRP NN Parse Tree DT NN Compositionality 57 Recurrent NN: CompositionalityRecurrent NN for Vector Space

57. NP IN NP DT NN PRP NN PP NP (S / ROOT) “rules” “meanings” Compositionality 58 Recurrent NN: CompositionalityRecurrent NN for Vector Space

58. Vector Space + Word Embeddings: Socher 59 Recurrent NN: CompositionalityRecurrent NN for Vector Space

59. Vector Space + Word Embeddings: Socher 60 Recurrent NN for Vector Space

60. Word Embeddings: Turian (2010) Turian, J., Ratinov, L., Bengio, Y. (2010). Word representations: A simple and general method for semi-supervised learning code & info: http://metaoptimize.com/projects/wordreprs/61

61. Word Embeddings: Turian (2010) Turian, J., Ratinov, L., Bengio, Y. (2010). Word representations: A simple and general method for semi-supervised learning code & info: http://metaoptimize.com/projects/wordreprs/ 62

62. Word Embeddings: Collobert & Weston (2011) Collobert, R., Weston, J., Bottou, L., Karlen, M., Kavukcuoglu, K., Kuksa, P. (2011) . Natural Language Processing (almost) from Scratch 63

63. Multi-embeddings: Stanford (2012) Eric H. Huang, Richard Socher, Christopher D. Manning, Andrew Y. Ng (2012)  Improving Word Representations via Global Context and Multiple Word Prototypes 64

64. Linguistic Regularities: Mikolov (2013) code & info: https://code.google.com/p/word2vec/ Mikolov, T., Yih, W., & Zweig, G. (2013). Linguistic Regularities in Continuous Space Word Representations 65

65. Word Embeddings for MT: Mikolov (2013) Mikolov, T., Le, V. L., Sutskever, I. (2013) .   Exploiting Similarities among Languages for Machine Translation 66

66. Word Embeddings for MT: Kiros (2014) 67

67. Recursive Deep Models & Sentiment: Socher (2013) Socher, R., Perelygin, A., Wu, J., Chuang, J.,Manning, C., Ng, A., Potts, C. (2013)   Recursive Deep Models for Semantic Compositionality Over a Sentiment Treebank. code & demo: http://nlp.stanford.edu/sentiment/index.html 68

68. Paragraph Vectors: Le & Mikolov (2014) Le, Q., Mikolov,. T. (2014) Distributed Representations of Sentences and Documents 69 • add context (sentence, paragraph, document) to word vectors during training ! Results on Stanford Sentiment   Treebank dataset:

69. Paragraph Vectors: Dai et al. (2014) 70

72. Global Vectors, GloVe: Stanford (2014) Pennington, P., Socher, R., Manning,. D.M. (2014).   GloVe: Global Vectors for Word Representation code & demo: http://nlp.stanford.edu/projects/glove/ vs results on the word analogy task “similar accuracy” 73

73. Dependency-based Embeddings: Levy & Goldberg (2014) Levy, O., Goldberg, Y. (2014). Dependency-Based Word Embeddings code & demo: https://levyomer.wordpress.com/2014/04/25/dependency-based-word- embeddings/ - Syntactic Dependency Context Australian scientist discovers star with telescope - Bag of Words (BoW) Context 0.3$ 0.4$ 0.5$ 0.6$ 0.7$ 0.8$ 0.9$ 1$ 0$ 0.1$ 0.2$ 0.3$ 0.4$ 0.5$ 0.6$ 0.7$ 0.8$ 0.9$ 1$ Precision$ Recall$ “Dependency-based embeddings have more functional similarities” 74

74. • LSTMS • Attention Wanna Play ? Recent breakthroughs 75

76. Wanna Play ? LSTM 77

78. Attention Gregor et al (2015) DRAW: A Recurrent Neural Network For Image Generation (arxiv) (code)

79. • Question-Answering Systems (&Memory) • Summarization • Text Generation • Dialogue Systems • Image Captioning & other multimodal tasks Wanna Play ? Recent breakthroughs 80

81. Wanna Play ? QA & Memory 82 • Memory Networks (Weston et al 2015) • Dynamic Memory Network (Kumar et al 2015) • Neural Turing Machine (Graves et al 2014) Facebook Metamind DeepMind Weston et al (2015) Memory Networks (arxiv)

82. QA & Memory 83 Yyer et al. (2014) A Neural Network for Factoid Question Answering over Paragraphs (paper)

83. Wanna Play ? QA & Memory 84 • Memory Networks (Weston et al 2015) • Dynamic Memory Network (Kumar et al 2015) • Neural Turing Machine (Graves et al 2014) Facebook Metamind DeepMind Zaremba & Sutskever (2015) Learning to Execute (arxiv)

84. Wanna Play ? QA & Memory 85 Babl Dataset

86. Wanna Play ? Text generation 87 Karpathy (2015), The Unreasonable Eﬀectiveness of Recurrent Neural Networks (blog)

87. Karpathy (2015), The Unreasonable Eﬀectiveness of Recurrent Neural Networks (blog)

89. Image-Text Embeddings 92 Socher et al (2013) Zero Shot Learning Through Cross-Modal Transfer (info)

90. Image-Captioning • Andrej Karpathy Li Fei-Fei , 2015.   Deep Visual-Semantic Alignments for Generating Image Descriptions (pdf) (info) (code) • Oriol Vinyals, Alexander Toshev, Samy Bengio, Dumitru Erhan , 2015. Show and Tell: A Neural Image Caption Generator (arxiv) • Kelvin Xu, Jimmy Ba, Ryan Kiros, Kyunghyun Cho, Aaron Courville, Ruslan Salakhutdinov, Richard Zemel, Yoshua Bengio, Show, Attend and Tell: Neural Image Caption Generation with Visual Attention (arxiv) (info) (code)

91. “A person riding a motorcycle on a dirt road.”??? Image-Captioning

92. “Two hockey players are ﬁghting over the puck.”??? Image-Captioning

93. “A stop sign is ﬂying in blue skies.” “A herd of elephants ﬂying in the blue skies.” Elman Mansimov, Emilio Parisotto, Jimmy Lei Ba, Ruslan Salakhutdinov, 2015. Generating Images from Captions with Attention (arxiv) (examples) Image-Captioning

94. • TensorFlow: Recently released library by Google.   http://tensorflow.org • Theano - CPU/GPU symbolic expression compiler in python (from LISA lab at University of Montreal). http://deeplearning.net/software/ theano/ • Caffe - Computer Vision oriented Deep Learning framework: caffe.berkeleyvision.org • Torch - Matlab-like environment for state-of-the-art machine learning algorithms in lua (from Ronan Collobert, Clement Farabet and Koray Kavukcuoglu) http://torch.ch/ • more info: http://deeplearning.net/software links/ Wanna Play ? General Deep Learning 97

95. • RNNLM (Mikolov)  http://rnnlm.org • NB-SVM  https://github.com/mesnilgr/nbsvm • Word2Vec (skipgrams/cbow)  https://code.google.com/p/word2vec/ (original)  http://radimrehurek.com/gensim/models/word2vec.html (python) • GloVe  http://nlp.stanford.edu/projects/glove/ (original)  https://github.com/maciejkula/glove-python (python) • Socher et al / Stanford RNN Sentiment code:  http://nlp.stanford.edu/sentiment/code.html • Deep Learning without Magic Tutorial:  http://nlp.stanford.edu/courses/NAACL2013/ Wanna Play ? NLP 98

96. Questions? roelof@kth.se www.csc.kth.se/~roelof/ 99 Code & Papers: Collaborative Open Computer Science .com @graphiﬁc