Switch branches/tags
Nothing to show
Find file History
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Type Name Latest commit message Commit time
..
Failed to load latest commit information.
Lecture1
Lecture2
Lecture3
Lecture4
Lecture5
Lecture6
.gitignore
README.md

README.md

Deep Learning for NLP - Lecture October 2015

This GIT repository accompanies the UKP lecture on Deep Learning for Natural Language Processing.

In contrast to other lectures, this lecture focuses on the usage of deep learning methods. As programming infrastructure we use Python in combination with Theano and Keras.

This lecture is structured into 6 parts. Each parts contains some recommended readings, which are supposed to be read before class. In class (video will be streamed and recorded) we will discuss the papers and provide some more background knowledge. With the start of the second lecture, each lecture will contain some practical exercise, in the most cases to implement a certain deep neural network to do a typical NLP task, for example Named Entity Recognition, Genre Classifcation of Sentiment Classification. The lecture is inspired by an engineering mindset: The beautiful math and complexity of the topic is sometimes neglected to provide instead an easy-to-understand and easy-to-use approach to use Deep Learning for NLP tasks (we use what works without providing a full background on every aspect).

At the end of the lecture you should be able to understand the most important aspect of deep learning for NLP and be able to programm and train your own deep neural networks.

In case of questions, feel free to approach Nils Reimers.

Recommended Readings on Deep Learning

The following is a short list with good introductions to different aspects of deep learning.

Lecture 1 - Introduction to Deep Learning

Monday, 5th October, 11am, Room B002

Video: https://youtu.be/AmG4jzmBZ88

Slides: pdf

Lecture-Content:

  • Overview of the lecture
  • What is Deep Learning? What is not Deep Learning?
  • Fundamental terminology
  • Feed Forward Neural Networks
  • Shallow vs deep neural networks

Recommended Readings

Lecture 2 - Introduction to Theano and Lasagne

Monday, 12th October, 11am, Room B002

Preparation before class:

Slides: pdf

Code: /Lecture2/code

  • The code uses Python 2.7. With Python 3, you might need to change the syntax accordingly

Video: https://youtu.be/BCwBl_55n7s

Lecture 3 - Word Embeddings and Deep Feed Forward Networks

Monday, 19th October, 11am, Room B002

Preparation before class:

  • Install Keras
  • Know the theory of word embeddings & word2vec
  • Watch from the CS224d Stanford Class the following videos:
    • Lecture 2
      • 00:00 - 21:30 - Introduction to word vectors via SVD
      • 21:30 - 28:00 - Hacks for word vector learning
      • 28:00 - 01:01:00- Problems with SVD, Introduction to word2vec (please watch at least this part)
        • From minute 38 to 53 he derives the optimization function of word2vec, feel free to skip this part
      • 1:01:00 - 1:13:00 - LSA vs. Skip-Gram vs. CBOW. Introduction of Glove
    • Lecture 3
      • 00:00 - 13:00 - How is word2vec trained, how are the word embeddings updated (please watch at least this part)
      • 13:00 - 20:00 - What is Skip-Gram, Negative Sampeling, CBOW (please watch at least this part)
      • 20:00 - 28:00 - How to evaluate word embeddings
      • 28:00 - 36:00 - How to improve the quality of word embeddings
      • 36:00 - 38:00 - Intrinsic evaluation of word embeddings
      • 38:00 - 41:00 - How to deal with ambiguous words?
      • 41:00 - 42:00 - Intrinsic evaluation of word embeddings
      • 42:00 - 50:45 - Using word embeddings and softmax for classification
      • 50:45 - 55:00 - Cross Entropy error
      • 55:00 - 1:08:00 - Should word embeddings be updated during classification?
  • The theory will not be introduced in class. But if you have questions regarding the theory / the videos, please ask them. We will discuss your questions / the videos in the beginning of Lecture 3
  • Get familiar with Theano and Lasagne, do the exercises from Lecture 2

Slides pdf

Video: https://youtu.be/MXHyIpv6RIg

  • There are a lot of parts where I use flipcharts and neither the audio nor the video captures this. Basically I present the Collobert et al., NLP almost from scratch, approach. An illustration of this drawing can be found here slides 41-43 (SENNA, Window Approach, Sentence Approach)

Exercises:

  • Try different hyperparameters, e.g. window size, number of hidden units, optimization function, activation functions
  • Take the NER Keras implementation (Lecture3/code/NER_Keras.py) and extend it with a Casing feature, i.e. the information if a word is all uppercase, all lowercase or initial uppercase. Hint: Your network needs two inputs, one for the word indices, one for the chasing information.

Lecture 4 - Autoencoders, Recursive Neural Networks, Dropout

Monday, 26th October, 11am, Room B002

Recommended Readings before class:

Slides: pdf

Video: https://youtu.be/FSKag11y8yI

Exercise: Have a look in the code directory. There you can find an example on the Brown corpus performing genre classifcation, one example on the 20 newsgroup dataset on topic classification and one example on autoencoders for the MNIST dataset.

Lecture 5 - Convolutional Neural Networks

Monday, 2nd November, 11am, Room B002

Recommended Readings:

Slides: pdf

Video: https://youtu.be/nzSPZyjGlWI

Lecture 6 - Recurrent models and LSTM-Model

Monday, 9h November, 11am, Room B002

Deep Learning for NLP - Lecture October 2015

This site can be access by the URL: www.deeplearning4nlp.com

This GIT repository accompanies the UKP lecture on Deep Learning for Natural Language Processing.

In contrast to other lectures, this lecture focuses on the usage of deep learning methods. As programming infrastructure we use Python in combination with Keras and Lasagne.

This lecture is structured into 6 parts. Each parts contains some recommended readings, which are supposed to be read before class. In class (video will be streamed and recorded) we will discuss the papers and provide some more background knowledge. With the start of the second lecture, each lecture will contain some practical exercise, in the most cases to implement a certain deep neural network to do a typical NLP task, for example Named Entity Recognition, Genre Classifcation of Sentiment Classification. The lecture is inspired by an engineering mindset: The beautiful math and complexity of the topic is sometimes neglected to provide instead an easy-to-understand and easy-to-use approach to use Deep Learning for NLP tasks (we use what works without providing a full background on every aspect).

At the end of the lecture you should be able to understand the most important aspect of deep learning for NLP and be able to programm and train your own deep neural networks.

In case of questions, feel free to approach Nils Reimers.

Recommended Readings on Deep Learning

The following is a short list with good introductions to different aspects of deep learning.

Lecture 1 - Introduction to Deep Learning

Monday, 5th October, 11am, Room B002

Video: https://youtu.be/AmG4jzmBZ88

Slides: pdf

Lecture-Content:

  • Overview of the lecture
  • What is Deep Learning? What is not Deep Learning?
  • Fundamental terminology
  • Feed Forward Neural Networks
  • Shallow vs deep neural networks

Recommended Readings

Lecture 2 - Introduction to Theano and Lasagne

Monday, 12th October, 11am, Room B002

Preparation before class:

Slides: pdf

Code: /Lecture2/code

  • The code uses Python 2.7. With Python 3, you might need to change the syntax accordingly

Video: https://youtu.be/BCwBl_55n7s

Lecture 3 - Word Embeddings and Deep Feed Forward Networks

Monday, 19th October, 11am, Room B002

Preparation before class:

  • Install Keras
  • Know the theory of word embeddings & word2vec
  • Watch from the CS224d Stanford Class the following videos:
    • Lecture 2
      • 00:00 - 21:30 - Introduction to word vectors via SVD
      • 21:30 - 28:00 - Hacks for word vector learning
      • 28:00 - 01:01:00- Problems with SVD, Introduction to word2vec (please watch at least this part)
        • From minute 38 to 53 he derives the optimization function of word2vec, feel free to skip this part
      • 1:01:00 - 1:13:00 - LSA vs. Skip-Gram vs. CBOW. Introduction of Glove
    • Lecture 3
      • 00:00 - 13:00 - How is word2vec trained, how are the word embeddings updated (please watch at least this part)
      • 13:00 - 20:00 - What is Skip-Gram, Negative Sampeling, CBOW (please watch at least this part)
      • 20:00 - 28:00 - How to evaluate word embeddings
      • 28:00 - 36:00 - How to improve the quality of word embeddings
      • 36:00 - 38:00 - Intrinsic evaluation of word embeddings
      • 38:00 - 41:00 - How to deal with ambiguous words?
      • 41:00 - 42:00 - Intrinsic evaluation of word embeddings
      • 42:00 - 50:45 - Using word embeddings and softmax for classification
      • 50:45 - 55:00 - Cross Entropy error
      • 55:00 - 1:08:00 - Should word embeddings be updated during classification?
  • The theory will not be introduced in class. But if you have questions regarding the theory / the videos, please ask them. We will discuss your questions / the videos in the beginning of Lecture 3
  • Get familiar with Theano and Lasagne, do the exercises from Lecture 2

Slides pdf

Video: https://youtu.be/MXHyIpv6RIg

  • There are a lot of parts where I use flipcharts and neither the audio nor the video captures this. Basically I present the Collobert et al., NLP almost from scratch, approach. An illustration of this drawing can be found here slides 41-43 (SENNA, Window Approach, Sentence Approach)

Exercises:

  • Try different hyperparameters, e.g. window size, number of hidden units, optimization function, activation functions
  • Take the NER Keras implementation (Lecture3/code/NER_Keras.py) and extend it with a Casing feature, i.e. the information if a word is all uppercase, all lowercase or initial uppercase. Hint: Your network needs two inputs, one for the word indices, one for the chasing information.

Lecture 4 - Autoencoders, Recursive Neural Networks, Dropout

Monday, 26th October, 11am, Room B002

Recommended Readings before class:

Slides: pdf

Video: https://youtu.be/FSKag11y8yI

Exercise: Have a look in the code directory. There you can find an example on the Brown corpus performing genre classifcation, one example on the 20 newsgroup dataset on topic classification and one example on autoencoders for the MNIST dataset.

Lecture 5 - Convolutional Neural Networks

Monday, 2nd November, 11am, Room B002

Recommended Readings:

Slides: pdf

Video: https://youtu.be/nzSPZyjGlWI

Lecture 6 - Recurrent models and LSTM-Model

Monday, 9h November, 11am, Room B002

Slides: pdf

Video: https://youtu.be/an9x3vXQYYQ