Deep Learning for Natural Language Processing: Word Embeddings
47 likes10,001 views
The document discusses the challenges and advancements in natural language processing (NLP), particularly focusing on deep learning techniques like word embeddings and recurrent neural networks. It outlines various applications of deep learning in NLP, such as sentiment analysis, machine translation, and text generation, while addressing issues related to language ambiguity, productivity, and cultural specificity. Additionally, it highlights the evolution and effectiveness of different neural network architectures for language understanding and modeling.
![[Karlgren 2014, NLP Sthlm Meetup]6](https://image.slidesharecdn.com/nlpguestlect15-151203142113-lva1-app6891/85/Deep-Learning-for-Natural-Language-Processing-Word-Embeddings-6-320.jpg)
![• 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](https://image.slidesharecdn.com/nlpguestlect15-151203142113-lva1-app6891/85/Deep-Learning-for-Natural-Language-Processing-Word-Embeddings-18-320.jpg)
![• 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](https://image.slidesharecdn.com/nlpguestlect15-151203142113-lva1-app6891/85/Deep-Learning-for-Natural-Language-Processing-Word-Embeddings-19-320.jpg)
![• 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](https://image.slidesharecdn.com/nlpguestlect15-151203142113-lva1-app6891/85/Deep-Learning-for-Natural-Language-Processing-Word-Embeddings-44-320.jpg)
![47
Recurrent Neural Networks
backpropagation through time
class based recurrent NN
[code (Mikolov’s RNNLM Toolkit) and more info: http://rnnlm.org/ ]](https://image.slidesharecdn.com/nlpguestlect15-151203142113-lva1-app6891/85/Deep-Learning-for-Natural-Language-Processing-Word-Embeddings-47-320.jpg)
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/ [email protected]
- 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
- 18. • 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
- 19. • 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
- 21. • Structure corresponds to meaning: Structure and Meaning 21
- 22. • Semantics • Syntax 22 NLP: what can we work with?
- 23. • Language models define probability distributions over (natural language) strings or sentences • Joint and Conditional Probability Language Model 23
- 24. • Language models define probability distributions over (natural language) strings or sentences Language Model 24
- 25. • Language models define probability distributions over (natural language) strings or sentences Language Model 25
- 26. 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
- 27. 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
- 28. 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
- 29. 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
- 30. 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
- 31. Distributional semantics Landauer and Dumais (1997), Turney and Pantel (2010), … 31
- 32. Distributional semantics Distributional meaning as co-occurrence vector: 32
- 33. 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
- 34. Word Embeddings: SocherVector Space Model adapted rom Bengio, “Representation Learning and Deep Learning”, July, 2012, UCLA In a perfect world: 34
- 35. 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
- 36. • 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
- 37. • 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
- 38. 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
- 39. 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
- 40. • How do we handle the compositionality of language in our models? 40 Compositionality
- 41. • 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
- 42. • How do we handle the compositionality of language in our models? • Option 1: Recurrent Neural Networks (RNN) 42 RNN 1: Recurrent Neural Networks
- 43. • 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
- 44. • 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
- 45. 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
- 46. 46 Recurrent Neural Networks backpropagation through time
- 47. 47 Recurrent Neural Networks backpropagation through time class based recurrent NN [code (Mikolov’s RNNLM Toolkit) and more info: http://rnnlm.org/ ]
- 48. • 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/
- 49. 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
- 50. Recursive Neural Tensor Network 50
- 51. • 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/
- 52. • 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/
- 53. • 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
- 54. • 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/
- 55. NP PP/IN NP DT NN PRP$ NN Parse Tree Recurrent NN for Vector Space 55
- 56. NP PP/IN NP DT NN PRP$ NN Parse Tree INDT NN PRP NN Compositionality 56 Recurrent NN: CompositionalityRecurrent NN for Vector Space
- 57. NP IN NP PRP NN Parse Tree DT NN Compositionality 57 Recurrent NN: CompositionalityRecurrent NN for Vector Space
- 58. NP IN NP DT NN PRP NN PP NP (S / ROOT) “rules” “meanings” Compositionality 58 Recurrent NN: CompositionalityRecurrent NN for Vector Space
- 59. Vector Space + Word Embeddings: Socher 59 Recurrent NN: CompositionalityRecurrent NN for Vector Space
- 60. Vector Space + Word Embeddings: Socher 60 Recurrent NN for Vector Space
- 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/61
- 62. 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
- 63. 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
- 64. 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
- 65. 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
- 66. Word Embeddings for MT: Mikolov (2013) Mikolov, T., Le, V. L., Sutskever, I. (2013) . Exploiting Similarities among Languages for Machine Translation 66
- 67. Word Embeddings for MT: Kiros (2014) 67
- 68. 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
- 69. 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:
- 70. Paragraph Vectors: Dai et al. (2014) 70
- 71. Paragraph Vectors: Dai et al. (2014) 71
- 72. Paragraph Vectors: Dai et al. (2014) 72
- 73. 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
- 74. 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
- 75. • LSTMS • Attention Wanna Play ? Recent breakthroughs 75
- 76. • LSTMS • Attention Wanna Play ? Recent breakthroughs 76
- 78. • LSTMS • Attention Wanna Play ? Recent breakthroughs 78
- 79. Attention Gregor et al (2015) DRAW: A Recurrent Neural Network For Image Generation (arxiv) (code)
- 80. • Question-Answering Systems (&Memory) • Summarization • Text Generation • Dialogue Systems • Image Captioning & other multimodal tasks Wanna Play ? Recent breakthroughs 80
- 81. • Question-Answering Systems (&Memory) • Summarization • Text Generation • Dialogue Systems • Image Captioning & other multimodal tasks Wanna Play ? Recent breakthroughs 81
- 82. 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)
- 83. QA & Memory 83 Yyer et al. (2014) A Neural Network for Factoid Question Answering over Paragraphs (paper)
- 84. 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)
- 85. Wanna Play ? QA & Memory 85 Babl Dataset
- 86. • Question-Answering Systems (&Memory) • Summarization • Text Generation • Dialogue Systems • Image Captioning & other multimodal tasks Wanna Play ? Recent breakthroughs 86
- 87. Wanna Play ? Text generation 87 Karpathy (2015), The Unreasonable Effectiveness of Recurrent Neural Networks (blog)
- 90. Karpathy (2015), The Unreasonable Effectiveness of Recurrent Neural Networks (blog)
- 91. • Question-Answering Systems (&Memory) • Summarization • Text Generation • Dialogue Systems • Image Captioning & other multimodal tasks Wanna Play ? Recent breakthroughs 91
- 92. Image-Text Embeddings 92 Socher et al (2013) Zero Shot Learning Through Cross-Modal Transfer (info)
- 93. 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)
- 94. “A person riding a motorcycle on a dirt road.”??? Image-Captioning
- 95. “Two hockey players are fighting over the puck.”??? Image-Captioning
- 96. “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
- 97. • 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
- 98. • 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
- 99. Questions? [email protected] www.csc.kth.se/~roelof/ 99 Code & Papers: Collaborative Open Computer Science .com @graphific