jaaack-wang/notes-for-stanford-cs224n-nlp-with-deep-learning

Notes for Stanford CS224N: Natural Language Processing with Deep Learning.

Notes for Stanford CS224N: Natural Language Processing with Deep Learning, a great course that I just discovered. You can also find the course videos on YouTube, which were recorded in Winter 2019 and contains 22 lecture videos. There are differences between the course slides found on the website (2021 version) and those used in the videos due to certain degrees of revisions. My notes are mainly based on the lecture videos, but also take the latest slides on the course website into accounts.

In each lecture folder, you will find the following materials: my notes, the lecture slide(s) and other course materials, such as course code or assignment. The assignment is left in a way as it is downloaded from the course website. In addition, I also placed the course notes written by Stanford students available on the course website in each lecture folder and denoted them by [Stanford notes] to distinguish from my notes, denoted as [Jack's notes]. I consciously made my notes as shareable as possible and I hope anyone who views this repository find them helpful.

You are also welcome to take a loot at my another repository dl-nlp-using-paddlenlp, which focuses the application of deep learning techniques in natural language processing using the state-of-the-art deep learning frameworks paddle and paddlenlp.

The "Table of Contents" below are structures of my notes taken for lectures that I finished.

Table of Contents

• Lecture 1-Intro and Word Vectors
• Lecture 2-Word Vectors and Word Senses
• Lecture 3-Neural Networks
• Lecture 4-Backpropagation
• Lecture 5-Dependency Parsing
• Lecture 6-Language Models and RNNs
• Lecture 7-Vanishing Gradients and Fancy RNNs
• Lecture 8-Translation, Seq2Seq, Attention
• Lecture 9-Practical Tips for Projects
• Lecture 10-ConvNets for NLP

Lecture 1-Intro and Word Vectors

• 1. Casual takeaways
• 2. WordNet
  • 2.1 Quick facts
  • 2.2 Accessing WordNet
  • 2.3 Limitations
• 3. Word vectors
  • 3.1 One-hot vectors
  • 3.2 Word vectors (embeddings)
  • 3.3 Word2vec algorithm basics
    • 3.3.1 Basic ideas
    • 3.3.2 Math behind
• 4. References
• 5. Appendix: gradients for loss function (2)

Lecture 2-Word Vectors and Word Senses

• 1. Casual takeaways
• 2. Word2vec
  • 2.1 Recap
  • 2.2 Two model variants
  • 2.3 The skip-gram model with negative sampling (HW2)
    • 2.3.1 Negative sampling
    • 2.3.2 Objective/loss functions
    • 2.3.3 Subsampling
• 3. Co-occurrence counts based method
  • 3.1 Co-occurrence matrix
  • 3.2 Dimensionality reduction
    • 3.2.1 Classic Method: Singular Value Decomposition
    • 3.2.2 Other tricks: scaling, pearson correlations etc.
  • 3.3 Count based method vs. direct prediction (neural network) method
• 4. GloVe: Global Vectors for Word Representation
  • 4.1 Basic idea
  • 4.2 Mathematical realization
• 5. Evaluation of word vectors
  • 5.1 Overview
  • 5.2 Intrinsic word vector evaluation
    • 5.2.1 Models
    • 5.2.2 Performances of different models
  • 5.3 Extrinsic word vector evaluation
  • 5.4 Factors that impact the performances
• 6. Challenges: word sense ambiguity
  • 6.1 Challenges
  • 6.2 Tentative solutions
• 7. References

Lecture 3-Neural Networks

• 1. Casual takeaways
• 2. Classification review and notation
  • 2.1 Inputs and labels
  • 2.2 Output models
  • 2.3 Objective/loss function and cross entropy
  • 2.4 Traditional ML classifier versus Neural Network Classifier
• 3. Neural Network Basics
  • 3.1 Neuron
  • 3.2 What an artificial neuron can do
  • 3.3 Matrix notation for a layer
• 4. Named Entity Recognition (NER)
  • 4.1 NER task
  • 4.2 Simple NER training methods
    • 4.2.1 Binary classification with averaged context words vectors
    • 4.2.2 Window classification using multi-class softmax classifier
    • 4.2.3 Binary classification with unnormalized scores using shallow neural network
  • 4.3 Challenges for NER
• 5. Gradients computation
  • 5.1 Partial derivatives and gradients
    • 5.1.1 Simple case: placed in a vector
    • 5.1.2 A more complex case: placed in a matrix -- Jacobian Matrix
  • 5.2 Multivariate calculus
    • 5.2.1 The chain rule
    • 5.2.2 Partial derivatives computation in Jacobian Matrix
    • 5.2.3 Shape convention
• 6. References

Lecture 4-Backpropagation

• 1. Casual takeaways
• 2. Derivative with regard to a weight matrix
  • 2.1 An shallow neural network example
  • 2.2 Let's derive
    • 2.2.1 With regard to W
    • 2.2.2 With regard to b
  • 2.3 Tips for deriving gradients
  • 2.4 Updating weights: gradient descent
  • 2.5 Transfer learning: using pre-trained parameters
• 3. Computation graph and backpropagation
  • 3.1 Overview
  • 3.2 An example
  • 3.3 Generalization of computation graph in neural networks
  • 3.4 Debugging
• 4. General tricks: regularization, nonlinearities, hyperparmeters etc.
  • 4.1 Regularization
  • 4.2 Vectorization
  • 4.3 Non-linearities: activation functions
  • 4.4 Parameter Initialization
  • 4.5 Optimizers
  • 4.6 Learning rate
• 5. References
• 6. Appendix
  • 6.1 Notations
  • 6.2 A quick and general solution
  • 6.3 Sigmoid function
  • 6.4 Softmax function

Lecture 5-Dependency Parsing

• 1. Casual takeaways
• 2. Syntactic structure: constituency and dependency
  • 2.1 Constituency
  • 2.2 Dependency
  • 2.3 Sources of ambiguity related to dependency
    • 2.3.1 Prepositional phrase attachment ambiguity
    • 2.3.2 PP attachment ambiguities multiply
    • 2.3.3 Coordination scope ambiguity
    • 2.3.4 Adjectival modifier ambiguity
    • 2.3.5 Verb phrase (VP) attachment ambiguity
• 3. Dependency grammar and dependency structure
  • 3.1 A bit history of dependency grammar/parsing
  • 3.2 Dependency grammar and its graphic representation
  • 3.3 Universal dependencies treebanks and annotated data
• 4. Dependency parser
  • 4.1 Dependency conditioning preferences
  • 4.2 General parsing rules
  • 4.3 Projectivity
  • 4.4 Methods of dependency parsing
• 5. Greedy transition-based parsing
  • 5.1 Overview
  • 5.2 Arc-standard transition-based parser
  • 5.3 ML transition-based parser
  • 5.4 Evaluation of accuracy
  • 5.5 Problems
• 6. Neural dependency parser
  • 6.1 General idea
  • 6.2 Preprocessing
  • 6.3 Model architecture
  • 6.4 Comparison
• 7. References

Lecture 6-Language Models and RNNs

• 1. Language Modeling
  • 1.1 Overview
  • 1.2 n-gram Language Models
  • 1.3 Fix-window Neural Language Model
  • 1.4 Evaluation: Perplexity
• 2. Recurrent Neural Networks
  • 2.1 Overview
    • 2.1.1 Basic architeture of RNN
    • 2.1.2 Applications
  • 2.2 RNN Language Model
    • 2.2.1 Example
    • 2.2.2 Pros and cons
    • 2.2.3 Training
• 3. Recap
• 4. References

Lecture 7-Vanishing Gradients and Fancy RNNs

• 1. Vanishing gradient
  • 1.1 Problem defined and cause
  • 1.2 Potential problems
  • 1.3 Possible consequences of Vanishing Gradient on RNN-LM in possible scenarios
• 2. Exploding gradient
  • 2.1 Problem defined and cause
  • 2.2 Potential problems
  • 2.3 Solutions: gradient clipping
• 3. Long Short-Term Memory (LSTM)
  • 3.1 Description
  • 3.2 Graphic representation
  • 3.3 LSTM success and replacement
• 4. Gated Recurrent Units (GRU)
  • 4.1 Description
  • 4.2 LSTM vs GRU
• 5. General solutions to vanishing/exploding gradient
  • 5.1 Vanishing/exploding gradient in NN
  • 5.2 Residual connections
  • 5.3 Dense connections
  • 5.4 Highway connections
• 6. Bidirectional RNNs
  • 6.1 Motivation
  • 6.2 Structure
  • 6.3 Restrictions
• 7. Multi-layer RNNs (Stacked)
  • 7.1 Description
  • 7.2 In practice
• 8. Summary
• 9. References

Lecture 8-Translation, Seq2Seq, Attention

• 1. Pre-Neural Machine Translation
  • 1.1 Problem defined
  • 1.2 Early 50s: rule-based
  • 1.3 1990s-2010s: statistical
• 2. Neural Machine Translation
  • 2.1 Problem defined
  • 2.2 Seq2seq Model
  • 2.3 Training
  • 2.4 Decoing
    • 2.4.1 Greedy decoding
    • 2.4.2 Exhaustive search decoding
    • 2.4.3 Beam search decoding
  • 2.5 Tradeoff of NMT
    • 2.5.1 Advantages
    • 2.5.2 Disadvantages
  • 2.6 Eluvation: BLEU
• 3. Attention
  • 3.1 Background
  • 3.2 Graphic represenation
  • 3.3 Equations
  • 3.4 Benifits
  • 3.5 Attention as a general DL technique
  • 3.6 Remaining problems
  • 3.7 Trend
• 4. References

Lecture 9-Practical Tips for Projects

• 1. Final project
  • 1.1 Default
  • 1.2 Two basic starting points of finding research topics
  • 1.3 Project types
  • 1.4 Interesting places to start
  • 1.5 Must-haves
  • 1.6 Where to find data
• 2. RNNs recap
  • 2.1 Recap of RNN
  • 2.2 Gated Recurrent Unit Recap
  • 2.3 Compare ungated and gated Unit
  • 2.4 Two most important variants of gated RNN
  • 2.5 LSTM
• 3. MT
  • 3.1 Problems
  • 3.2 Solutions
  • 3.3 MT evaluation
  • 3.4 BLEU
• 4. Steps of Working on a project
• 5. References

Lecture 10-ConvNets for NLP

• 1. Convolutional Nerual Netwrok (CNN)
  • 1.1 Overview
  • 1.2 2D example
  • 1.3 1D example
    • 1.3.1 With padding
    • 1.3.2 With multiple filters
    • 1.3.3 With Stride = 2
    • 1.3.4 k-max pooling
    • 1.3.5 Dilated CNN
    • 1.3.6 PyTorch implementation (example)
• 2. Yoon Kim (2014)
  • 2.1 Single layer CNN for sentence classification
  • 2.2 Hyperparameters
  • 2.3 Model Variants and results
• 3. Model comparisons, related techniques and applications
  • 3.1 Comparisons: Bag of Vectors, Window Model, CNNs, RNNs
  • 3.2 Techniques
    • 3.2.1 Batch Normalization
    • 3.2.2 1 x 1 Convolutions
  • 3.3 Application
    • 3.3.1 Translation
    • 3.3.2 POS tagging
    • 3.3.3 Character-Aware Neural Language Models

notes-for-stanford-cs224n-nlp-with-deep-learning's Issues

Minor formula notation issue for lecture 1 note

Dear Jack,
Thank you so much for this wonderful knowledge share. My encounters with you have made me better in this field!

There are very minor notation issues for lecture 1 note. Others are just perfect!

1. formula (1), should be given the centor word $w_t$, and the likelihood should be $$Likelihood = L(\theta) = \prod_{t}^{T} \prod_{\substack{-m \leq j \leq m \\ j \neq 0}} P(w_{t+j}|{w_{t}; \theta})$$

The reason to change to $$j \neq0 $$, rather not $$j \neq m$$ is when $j$ is 0, it means the center word $w_t$ itself. Similar issue applies to other formula.

2. For formula(3), the wording for prediction function should be : function of predicting context word $v_{o}$ , given the center word $v_{c}$ and the vocabulary $V$. Similarly, formula (5), it should be $$J(\mathbf{u_{o} \;|\; v_{c}}) $$
3. For Partial derivative with regard to $v_c$ section, step 3
   $$\frac{\partial}{\partial \mathbf{v_c}}exp(\mathbf{u_{x}^{T} v_{c}}) = exp(\mathbf{u_{x}^{T} v_{c}}) \cdot \frac{\partial}{\partial \mathbf{v_c}}\mathbf{u_{x}^{T} v_{c}} = exp(\mathbf{u_{x}^{T} v_{c}}) \cdot \mathbf{u_{x}}\tag{Step 3}$$. I think $u_x$ here should indicate a vector $$\mathbf{u_{x}}$$.

Let me know if my understanding is correct or not. Again, thank you for generously sharing your time with community!