Deep Learning for Information Retrieval

@graphificRoelof Pieters

Guest  Lecture:  Deep  Learning  for  Informa8on  Retrieval28  April  2015

www.csc.kth.se/~roelof/ [email protected]

[email protected]

Gve Systems Graph Technologies R&D

DD2476 Search Engines and Information Retrieval Systemshttps://www.kth.se/social/course/DD2476/

slides  online  at  h4p://www.slideshare.net/roelofp/deep-­‐learning-­‐for-­‐informa=on-­‐retrieval  

2

About Me • (-10y) CS dropout (Amsterdam Technical Univ.)• (2y) Msc Social Anthropology, Stockholm

University• Current: PhD candidate at KTH/CSC with focus

on:• Deep Learning for Natural Language

Processing (Distributed Semantics) • Graph-based approaches for Knowledge

Representation• Multi-modal models

• Current: Data Science Consultant at Graph Technologies RD & Gve-Systems• Recommender Systems• Deep Learning• Realtime Graph-based Search Engines

3

Information Retrieval (IR)

- Hedvig Kjellström, lecture 1

4

Data landscape is changing

1. Amount of digital data is growing at increasing rate (IOT, digitalization, wearables, phones/tablets)

2. Data types are shifting as well:

1. from text to audio-visual

2. from professional to personal/social (social media)

3. from semi-structured to unstructured

[Jussi Karlgren, NLP Sthlm Meetup 2014]

6

Data landscape is changing

Triple V’s of Big Data:1. Volume2. Velocity3. Variety

7

Making sense of DataTypical ML Regression

8

Making sense of DataNeural NetTypical ML Regression

Degrees of Complexity

9perceptron demo

Neural Net

10

(figure from Lior Rokach, Ben-Gurion University)

Neural Net

11

(figure from Lior Rokach, Ben-Gurion University)

Neural Net

12

(figure from Lior Rokach, Ben-Gurion University)

Neural Net

13

(figure from Lior Rokach, Ben-Gurion University)

Neural Net

14

(figure from Lior Rokach, Ben-Gurion University)

Neural Net

15

multilayer nn demo

Deep Learning ??

16

Deep Learning ??

17

• Learning multiple layers• “Back propagation”• Can “theoretically” learn any function!

Prior to 2006:• Very slow and inefficient• SVMs, random forests, etc. SOTA

18

2006+: the 3 Deep Learning Conspirators

19

20

— Andrew Ng

“I’ve worked all my life in Machine Learning, and I’ve

never seen one algorithm knock over benchmarks like Deep

Learning”

Deep Learning: Why?

21

Different Levels of Abstraction

22

Hierarchical Learning• Natural progression

from low level to high level structure as seen in natural complexity

Different Levels of AbstractionFeature Representation

23

Hierarchical Learning• Natural progression

from low level to high level structure as seen in natural complexity

• Easier to monitor what is being learnt and to guide the machine to better subspaces

Different Levels of AbstractionFeature Representation

24

Hierarchical Learning• Natural progression

from low level to high level structure as seen in natural complexity

• Easier to monitor what is being learnt and to guide the machine to better subspaces

• A good lower level representation can be used for many distinct tasks

Different Levels of AbstractionFeature Representation

25

Hierarchical Learning• Natural progression

from low level to high level structure as seen in natural complexity

Different Levels of AbstractionFeature Representation

2626

• Easier to monitor what is being learnt and to guide the machine to better subspaces

• A good lower level representation can be used for many distinct tasks

Classic Deep Architecture

Input layer

Hidden layers

Output layer

27

Modern Deep Architecture

Input layer

Hidden layers

Output layer

movie time:http://www.cs.toronto.edu/~hinton/adi/index.htm

28

[Kudos to Richard Socher, for this eloquent summary :) ]

• Manually designed features are often over-specified, incomplete and take a long time to design and validate

• Learned Features are easy to adapt, fast to learn

• Deep learning provides a very flexible, (almost?) universal, learnable framework for representing world, visual and linguistic information.

• Deep learning can learn unsupervised (from raw text/audio/images/whatever content) and supervised (with specific labels like positive/negative)

Why Deep Learning ?

29

Word Embeddings

30

31

What about NLP ?

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:

• 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 ?

Language 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 …] ANDStore [0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 …] = 0 !

32

Language Representation

33

- Johan Boye, lecture 2

Term-document matrix = Sparse!

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 models34

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)35

Distributional semantics

Landauer and Dumais (1997), Turney and Pantel (2010), …36

Distributional semanticsDistributional meaning as co-occurrence vector:

37

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 =

38

Word Embeddings: SocherVector Space Model

adapted rom Bengio, “Representation Learning and Deep Learning”, July, 2012, UCLA

In a perfect world:

39

Word Embeddings: SocherVector Space Model

adapted rom Bengio, “Representation Learning and Deep Learning”, July, 2012, UCLA

In a perfect world:

the country of my birththe place where I was born

40

• 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?

41

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 birththe place where I was born ?

…

42

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”

43

• How do we handle the compositionality of language in our models?

44

Compositionality

• How do we handle the compositionality of language in our models?

• Recursion :the same operator (same parameters) is applied repeatedly on different components

45

Compositionality

• How do we handle the compositionality of language in our models?

• Option 1: Recurrent Neural Networks (RNN)

46

RNN 1: Recurrent Neural Networks

(we ignore recurrent NN’s for this talk)

• How do we handle the compositionality of language in our models?

• Option 2: Recursive Neural Networks (also sometimes called RNN)

47

RNN 2: Recursive Neural Networks

Recursive Neural Tensor Network

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

NP

PP/IN

NP

DT NN PRP$ NN

Parse TreeRecurrent NN for Vector Space

50

NP

PP/IN

NP

DT NN PRP$ NN

Parse Tree

INDT NN PRP NN

Compositionality

51

Recurrent NN: CompositionalityRecurrent NN for Vector Space

NP

IN

NP

PRP NN

Parse Tree

DT NN

Compositionality

52

Recurrent NN: CompositionalityRecurrent NN for Vector Space

NP

IN

NP

DT NN PRP NN

PP

NP (S / ROOT)

“rules” “meanings”

Compositionality

53

Recurrent NN: CompositionalityRecurrent NN for Vector Space

Vector Space + Word Embeddings: Socher

54

Recurrent NN: CompositionalityRecurrent NN for Vector Space

Vector Space + Word Embeddings: Socher

55

Recurrent NN for Vector Space

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/ 56

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/ 57

Word Embeddings: Demo

Word Embeddings: Collobert & Weston (2011)

Collobert, R., Weston, J., Bottou, L., Karlen, M., Kavukcuoglu, K., Kuksa, P. (2011) . Natural Language Processing (almost) from Scratch

59

Polysemous-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

60

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

61

Word Embeddings for MT: Mikolov (2013)

Mikolov, T., Le, V. L., Sutskever, I. (2013) . Exploiting Similarities among Languages for Machine Translation

62

Word Embeddings for MT: Kiros (2014)

Kiros, R., Zemel, R. S., Salakhutdinov, R. (2014) . A Multiplicative Model for Learning Distributed Text-Based Attribute Representations

63

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.html64

Paragraph Vectors: Le & Mikolov (2014)

Le, Q., Mikolov,. T. (2014) Distributed Representations of Sentences and Documents65

• add context (sentence, paragraph, document) to word vectors during training

!

Results on Stanford Sentiment Treebank dataset:

Paragraph Vectors: Dai et al. (2014)

Dai, A., Olah,. C., Le, Q., Corrado, G. (2014) Document Embedding with Paragraph Vectors66

Paragraph Vectors: Dai et al. (2014)

Dai, A., Olah,. C., Le, Q., Corrado, G. (2014) Document Embedding with Paragraph Vectors67

Paragraph Vectors: Dai et al. (2014)

Dai, A., Olah,. C., Le, Q., Corrado, G. (2014) Document Embedding with Paragraph Vectors68

Nearest neighbours to the machine learning paper “Distributed Representations of Sentences and Documents” in arXiv.

Joint Image-Word Embeddings

69

1. Multimodal representation learning

2. Generating descriptions of images

3. Ranking images and captions (“image-sentence ranking”)

Some Current Approaches

70

Bags of Visual Words

71

Source credit : K. Grauman, B. Leibe

Bags of Visual Words (Sivic & Zisserman 2003)

standard BoW issues however

What we get:

But we want:• visual word order/relations• location• scale/viewpoint invariance• …

72

Zero-shot Learning

• skip-gram text model on wikipedia corpus of 5.7 million documents (5.4 billion words) - approach from (Mikolov et al. ICLR 2013)

73

Frome, A., Corrado, G.S., Shlens, J., Bengio, S., Dean, J., Mikolov, T., Ranzato, M.A. (2013) Devise: A deep visual-semantic embedding model

DeViSE model

Encoder: A deep convolutional network (CNN) and long short-term memory recurrent network (LSTM) for learning a joint image-sentence embedding. Decoder: A new neural language model that combines structure and content vectors for generating words one at a time in sequence.

Encoder-Decoder pipeline

74

Kiros, R., Salakhutdinov, R., Zemerl, R. S. (2014) Unifying Visual-Semantic Embeddings with Multimodal Neural Language Models

(Kiros et al 2014)

• captures Multimodal linguistic regularities

Encoder-Decoder pipeline

75

• captures Multimodal linguistic regularities

Encoder-Decoder pipeline

76

(PCA projection of (300-dimensional) word and image representations)

77

Vinyals, O., Toshev, A., Bengio, S., Erhan. D. (2015) Show and Tell: A Neural Image Caption Generator

Joint Visual-Semantic embedding

Karpathy, A., Fei Fei, L. (2015) Deep Visual-Semantic Alignments for Generating Image Descriptions

CNN+LSTM

CNN+RNN

78

Karpathy, A., Fei Fei, L. (2015) Deep Visual-Semantic Alignments for Generating Image Descriptions

Joint Visual-Semantic embedding

79

Joint Visual-Semantic embedding

Karpathy, A., Fei Fei, L. (2015) Deep Visual-Semantic Alignments for Generating Image Descriptions

80

Joint Visual-Semantic embedding

Karpathy, A., Fei Fei, L. (2015) Deep Visual-Semantic Alignments for Generating Image Descriptions

Joint Visual-Semantic embedding

81

Karpathy, A., Fei Fei, L. (2015) Deep Visual-Semantic Alignments for Generating Image Descriptions

demo

Any Questions?

Download example code samples fromhttps://github.com/graphific/DL-Meetup-intro

83

git clone --recursive https://github.com/graphific/DL-Meetup-intro.git

Wanna Play ? Code!

(more at http://deeplearning.net/ )

• Theano - CPU/GPU symbolic expression compiler in python (from LISA lab at University of Montreal). http://deeplearning.net/software/theano/

• Pylearn2 - library designed to make machine learning research easy. http://deeplearning.net/software/pylearn2/

• 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 ?

Wanna Play ? General Deep Learning

84

• 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)

• GloVehttp://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

85

• cuda-convnet2 (Alex Krizhevsky, Toronto) (c++/CUDA, optimized for GTX 580) https://code.google.com/p/cuda-convnet2/

• Caffe (Berkeley) (Cuda/OpenCL, Theano, Python)http://caffe.berkeleyvision.org/

• OverFeat (NYU) http://cilvr.nyu.edu/doku.php?id=code:start

Wanna Play ? Computer Vision

86

87

Impact on Computer Vision

88

Impact on Computer Vision

(from Clarifai)89

Impact on Audio ProcessingSpeech Recognition

90

Impact on Audio ProcessingTIMIT Speech Recognition

(from: Clarifai)91

C&W 2011

Impact on Natural Language Processing

Pos: Toutanova et al. 2003)Ner: Ando & Zhang 2005

C&W 2011

92

Impact on Natural Language Processing

Named Entity Recognition:

93