Natural Language Processing 2022

Intro (Demos)

A few cool things we'll (probably) learn more about in this class:

• Extracting relevant keywords from text
• Translating text from (say) Slovak to English
• Answering questions, given a specific paragraph of text, an image or even a table
• Automatically autocomplete comprehensible text
• Generate interesting text (like song lyrics)
• Summarizing text (i.e. give it a paragraph and get a single sentence back)

Lecture I: Text Processing

Discussed material:

• Basic Text Processing (slides)
  • Regular Expressions
  • Tokenization
  • Text normalization
• Basically the first part of SLP Chapter II

Supplementary resources:

• Eliza Bot Demo: one of the most famous use cases of regular expressions. It is really worth trying out -- you may end up having some surprisingly good conversations.
• Unix for poets: a nice 25 pages worth of examples on how to process text on the Unix command line. Here is a shorter version by the authors of the SLP book.
• Scriptio continua: the reason why English also nearly ended up without word and sentence separators.

Lecture II: Edit Distance

Discussed material:

• Edit Distance (slides)
  • Edit Distance
  • Weighted Edit Distance
  • Alignment
  • The last part of SLP Chapter II
  • (we did not discuss the bio applications ...)

Supplementary resources:

• subsync: a tool for automatically synchronizing subtitles with video (a nice use-case of using alignment in a not-so-ordinary context)
• Seam carving: although not necessarily NLP related, seam carving is a very nice real-world example of how dynamic programming can still be a useful tool.

Lecture III: Language Modeling with N-grams

Discussed material:

• Language modeling
  • Estimating N-gram probabilities
  • Perplexity and Language Model Evaluation
  • Dealing with zeros
  • Smoothing, backoff and interpolation
  • Most of SLP Chapter III

Supplementary resources:

• Google Ngram Viewer -- a nice way of visualizing the rate of use of n-grams in books written in various languages. Check out this quick example for instance. Note that the dataset on which the Ngrams are computed plays a big role -- here is a similar example for coronavirus.
• kenlm -- an open-source language modeling toolkit. Probably best in class when it comes to speed and memory efficiency.

Lecture IV: Language Modeling and Word Embeddings

"From N-grams to Word2Vec"

Discussed material:

• Language modeling and word embeddings
  • Shortcomings of n-gram language models
  • Neural language model
  • Word2Vec
  • CBOW + Skip-Gram
  • Visualization of (vector) word spaces

Supplementary resources:

• wevi -- visualizes what actually happens when word2vec-style embeddings are being learned in by a (relatively simple) neural network. The great thing about this demo is that it is implemented completely using in-browser technologies (i.e. JavaScript) and hence you get to "see the network close-up".
• Embedding Projector -- a nice way of visualizing distributed representations obtained using Word2Vec in 2D and 3D from the TensorFlow project
• Semantic Space Surfer -- a bit of a fun spin on word embeddings. While normally we'd like the word embeddings to work "intuitively" (i.e. king should be to man as queen is to woman), we know it's not always like that. In this quick game you're forced to "think like a word embedding" and pick the word that the pre-trained embeddings would have chosen. Not only is this quite fun, it'll also help you deepen your intuition around what's actually going on with word embeddings.

Lecture V: Language Modeling with RNNs

"From Word2Vec to Recurrent Networks"

Discussed material:

• Language Modeling and RNNs Part 1
  • Recurrent Neural Networks (RNNs)
  • Back Propagation Through Time (BPTT)
• Language Modeling and RNNs Part 2
  • Simple RNNs
  • Vanishing and Exploding Gradient
  • Long Short Term Memory (LSTM)
  • Gated Recurrent Unit (GRU)

Supplementary resources:

• SCIgen - An Automatic CS Paper Generator -- exactly what it sounds like.
• The Unreasonable Effectiveness of Recurrent Neural Networks -- at this point a classical blog post in Deep Learning blogosphere by Andrej Karpathy on what RNNs are capable of. Still worth checking out.
• The unreasonable effectiveness of Character-level Language Models -- a quick "setting-the-record-straight" response to the previous blog post by Yoav Goldberg. Note that it is not entirely critical, as its subtitle is (and why RNNs are still cool). Definitely worth a read.
• Unsupervised Sentiment Neuron -- a blogoduction (introduction-via-blog) of OpenAI for their work on training character-level language models on Amazon reviews, which happen to pick up the notion of "sentiment" on their own.
• Understanding LSTM Networks -- another classic which does a great job introducing LSTMs from the ground up with nicely done visualizations using the "circuit board" metaphor.

Lecture VI: Text Classification

Discussed material:

• Text Classification and Naive Bayes
  • Text Classification task
  • Bag of Words representation
  • Naive Bayes classifier
  • Classification metrics: accuracy, precision, recall, F score
  • Micro vs Macro averaging
• Sentiment Analysis
  • Sentiment Analysis Task
  • Sentiment Analysis using Naive Bayes classification
  • Lexicon-based approaches

Supplementary resources:

• Trump Tweet Bot: being able to estimate sentiment of text can have real-world implications: by assessing what a leader of a large country thinks about a publicly traded company and buying/selling its shares as a result, you can end up with a fairly interesting investment portfolio!
• Bag of Tricks for Efficient Text Classification: the paper which has introduced the fastText classifier as a simple competitive baseline (compared to deep learning models), with much more effective training procedure.
• Enriching Word Vectors with Subword Information: the paper which presents and interesting approach for dealing with Out Of Vocabulary words that is already present in fastText -- combining word vectors with embeddings of character n-grams.

Lecture VII: Tranformers and BERT

Discussed material:

The Illustrated Transformer

• The limits of attention
• Values, Keys and Queries
• Multi Head Attention
• Tricks of the Transformer architecture (i.e. positional embeddings)

Contextual representations via BERT

• Word vectors and issues with using them in context-free manner
• Representations with Language Models
• ELMo: Deep Contextual Word Embeddings
• Transformers and Self-Attention
• Masked Language Models

Supplementary resources:

Transformers from scratch:

If you ever happen to reimplement a Transformer from scratch, I strongly recommend following this tutorial. It assumes the understanding of the basic neural network concepts as well as backpropagation and some working knowledge of PyTorch, but other than that it is a self-contained, zero-to-hero guide that will walk you through the whole process from the smallest building blocks (like self-attention) to going big with the architecture.

The Illustrated BERT:

Although the presentation we discussed is very informative, a visual presentation is usually even more impactful. This is one of the best ones that you can find online.

Byte Pair Encoding:

Dealing with out of vocabulary words out-of-the-box was one of the big improvements BERT-style pre-trained models made very popular. This article goes into greater detail on how does that happen, what are the limits of this method and how one may go about fixing them.

(I actually recommend you read through the whole article -- it's a very nice introduction to the concept of Attention and sequence-to-sequence tasks in general)

The Dark Secrets of BERT:

It turns out BERT learns interesting things during training. Part of it may be due to its use of self-attention but as this article (and associated paper) shows, there may be some black magic going on.