Michael Elhadad

Topics in Natural Language Processing (202-2-5381) Fall 2019

Meets: Sun 10-12 Bdg 34 Room 114


  1. 27 Jan 19: Registration for HW3 grading - email me to reserve your preferred slot.
  2. 30 Dec 18: Quizz 09 - Lecture 10: Syntactic Analysis (2)
  3. 24 Dec 18: HW3 is posted - due date Thu 10 Jan Midnight.
  4. 23 Dec 18: Quizz 08 - Lecture 9: Syntactic Analysis (1)
  5. 21 Dec 18: Registration for HW2 grading - email me to reserve your preferred slot.
  6. 10 Dec 18: There will be no class on Sunday 16 Dec
  7. 02 Dec 18: Quizz 07 - Lecture 8: Sequence Classification and Structured Prediction
  8. 29 Nov 18: HW2 is posted - due date Tue 18 Dec Midnight.
  9. 28 Nov 18: Registration for HW1 grading - email me to reserve your preferred slot.
  10. 25 Nov 18: Quizz 06 - Lecture 7: Classification for NLP
    Extension for HW1 submission to Wed 28 Nov Evening
  11. 18 Nov 18: Quizz 05 - Lecture 6: Basic Statistical Concepts for ML in NLP
  12. 09 Nov 18: Quizz 04 - Lecture 5: Parts of Speech Tagging
  13. 09 Nov 18: HW1 is posted - due date Mon 26 Nov 2018 Midnight.
  14. 04 Nov 18: Quizz 03 - Lecture 4: Word Embeddings
  15. 28 Oct 18: Quizz 02 - Lecture 3: Intro to Deep Learning and Neural Language Models
  16. 21 Oct 18: Quizz 01 - Lecture 2: Language Models
  17. 14 Oct 18: Welcome to NLP 19 - Lecture 1


  1. General Intro to NLP - Linguistic Concepts
  2. Language Modeling
  3. Deep Learning Intro and Neural Language Models
  4. Word Embeddings
  5. Parts of Speech Tagging
  6. Basic Statistical Concepts for ML in NLP
  7. Classification
  8. Sequence Classification
  9. Syntactic Analysis (1/2): PCFGs
  10. Syntactic Analysis (2/2): Dependency Parsing


The course is an introduction to Natural Language Processing. The main objective of the course is to learn how to develop practical computer systems capable of performing intelligent tasks on natural language: analyze, understand and generate written text. This task requires learning material from several fields: linguistics, machine learning and statistical analysis, and core natural language techniques.
  1. Acquire basic understanding of linguistic concepts and natural language complexity: variability (the possibility to express the same meaning in many different ways) and ambiguity (the fact that a single expression can refer to many different meanings in different contexts); levels of linguistic description (word, sentence, text; morphology, syntax, semantics, pragmatics). Schools of linguistic analysis (functional, distributional, Chomskyan); Empirical methods in Linguistics; Lexical semantics; Syntactic description; Natural language semantics issues.
  2. Acquire basic understanding of machine learning techniques as applied to text: supervised vs. unsupervised methods; training vs. testing; classification; regression; distributions, KL-divergence; Bayesian methods; Support Vector Machines; Perceptron; Deep Learning methods in NLP; RNNs, LSTMs, sequence to sequence models and attention.
  3. Natural language processing techniques: word and sentence tokenization; parts of speech tagging; lemmatization and morphological analysis; chunking; named entity recognition; language models; probabilistic context free grammars; probabilistic dependency grammars; parsing accuracy metrics; treebank analysis; text simplification; paraphrase detection; summarization; text generation; topic modelling.
Topics covered in class include:
  1. Descriptive linguistic models
  2. Language Models -- Statistical Models of Unseen Data (n-gram, smoothing, recurrent neural networks language models)
  3. Language Models and deep learning -- word embeddings, continuous representations, neural networks
  4. Parts of speech tagging, morphology, non categorical phenomena in tagging
  5. Information Extraction / Named Entity Recognition
  6. Using Machine Learning Tools: Classification, Sequence Labeling / Supervised Methods / SVM. CRF, Perceptron, Logistic Regression
  7. Bayesian Statistics, generative models, topic models, LDA
  8. Syntactic descriptions: Parsing sentence, why, how, PCFGs, Dependency Parsing
  9. Text Summarization

Lecture Notes
  1. 14 Oct 18: General Intro to NLP - Linguistic Concepts

    Things to do:

    1. Find a way to estimate how many words exist in English. In Hebrew. What method did you use? What definition of word did you use? (Think derivation vs. inflection)
    2. Experiment with Google Translate: find ways to make Google Translate "fail dramatically" (generate very wrong translations). Explain your method and collect your observations. Document attempts you made that did NOT make Google Translate fail. (Think variability and ambiguity; Think syntactic complexity; think lexical sparsity, unknown words).
    3. Think of reasons why natural languages have evolved to become ambiguous (Think: what is the communicative function of language; who pays the cost for linguistic complexity and who benefits from it; is ambiguity created willingly or unconsciously?)

  2. 21 Oct 18: Language Modeling
    1. * N-gram Language Modeling Chapter 3 from Speech and Language Processing (SPL3), Jurafsky and Martin, 3rd Ed, 2016. (Focus on 3.1-3.4). If you prefer - the same material covered in slides format.
    2. * Peter Norvig: How to Write a Spelling Corrector (2007).

      This is a toy spelling corrector illustrating the statistical NLP method (probability theory, dealing with large collections of text, learning language models, evaluation methods). Read an extended version of the material with more applications (word segmentation, n-grams, smoothing, more on bag of words, secret code decipher): How to Do Things with Words (Use this local copy adapted to Python 3 / html version and the support files: big.txt, count_1w.txt, count_2w.txt).

    3. Spelling Correction and the Noisy Channel, Appendix B from SPL3, B.1-B.2. This covers similar material as Norvig's piece above in a more formal manner.
    4. Things to do:
      1. * Read about Probability axioms and in more details in the notes from Fredrik Engstrom:
      2. Read about Edit Distance and in more details, a review of minimum edit distance algorithms using dynamic programming from Dan Jurafsky.
      3. * Install Python: I recommend installing the Anaconda distribution (choose the Python 3.7 version). (Note: Many of the code samples you will see are written in Python 2 - which is not exactly compatible with Python 3 - the main annoying difference is that in Python 2 you can write: print x -- in Python 3 it must be print(x). Python3 comes with a utility 2to3 which converts most of the differences between Python 2 and Python 3.]

        The Anaconda distribution includes a large set of Python packages ready to use that we will find useful. (630MB download, 3GB disk space needed.) In particular, Anaconda includes the nltk, pandas, numpy, scipy and scikit-learn packages.

      4. Execute Norvig's spell checker for English (you will neeed the Python code from the article and the large file of text used for training big.txt).
      5. How many tokens are there in big.txt? How many distinct tokens? What are the 10 most frequent words in big.txt?
      6. In a very large corpus (discussed in the ngram piece quoted below), the following data is reported:
        The 10 most common types cover almost 1/3 of the tokens, the top 1,000 cover just over 2/3.
        What do you observe on the much smaller big.txt corpus?
      7. Read more from Norvig's piece on ngrams.
      8. Execute the word segmentation example from Norvig's ngram chapter (code in this notebook).

        Note the very useful definition of the @memo decorator in this example, which is an excellent method to implement dynamic programming algorithms in Python. From Python Syntax and Semantics:

        A Python decorator is any callable Python object that is used to modify a function, method or class definition. A decorator is passed the original object being defined and returns a modified object, which is then bound to the name in the definition. Python decorators were inspired in part by Java annotations, and have a similar syntax; the decorator syntax is pure syntactic sugar, using @ as the keyword:
        def menu_item():
        is equivalent to:
        def menu_item():
        menu_item = viking_chorus(menu_item)
      9. This corpus includes a list of about 40,000 pairs of words (error, correction). It is too small to train a direct spell checker that would map word to word. Propose a way to learn a useful error model (better than the one used in Norvig's code) using this corpus. Hint: look at the model of weighted edit distance presented in Jurafsky's lecture cited above.

  3. 28 Oct 2018: Deep Learning Intro and Neural Language Models

    Things to do:

    1. * Learn Python if you don't know it (About 4 hours)
    2. Install a good Python environment (About 2 hours) The default environment is pyCharm (the free Community Edition is good for our needs 127MB download). If you already have experienced with it, VSCode also has very good support for Python.
    3. * Follow the NumPy tutorial to learn how to manipulate arrays and tensors efficiently in Python. (2 hours) Check your knowledge with numpy exercises (1 hour).
    4. * Install Pytorch in your environment (1 hour):
      Install Anaconda for Python 3.7.
      Install PyTorch: follow instructions from pytorch.org and download the stable Python 3.7.
      If you have an nVidia GPU on your machine, select the appropriate CUDA version that you have installed.
      (Usually, the command to execute is "conda install pytorch -c pytorch").
    5. * Explore the PyTorch tutorials for text - starting from Deep Learning NLP Tutorial - cover the first 3 segments (3 hours).

  4. 04 Nov 18 Word Embeddings

    1. * View Chris Manning's Lecture on Word2Vec April 2017 (1h10)
    2. * Read CS 224D: Deep Learning for NLP1 1 - Lecture Notes: Part I by Richard Socher / the corresponding slides - Pre-trained word embeddings.
    3. Execute the following PyTorch tutorials on dense word representations in deep learning:

    Further Reading::

    1. Read Vector Semantics, Chapter 6 of Speech and Language Processing 3rd Ed, Jurafsky and Martin 2018.
    2. Read Efficient Estimation of Word Representations in Vector Space from Mikolov et al (2013) and the Word2vec site.
    3. Read Deep contextualized word representations, Peters et al 2018 (ELMO)
    4. Read BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding, Devlin et al 2018 (BERT).
    5. Read TensorFlow's tutorial on Word Embeddings.
    6. Read Word2Vec Explained by Yoav Goldberg and Omer Levy, 2014.
    7. Read Neural Word Embedding as Implicit Matrix Factorization by Levy and Goldberg, 2014.
    8. Read king - man + woman is queen; but why? by Piotr Migdal, Jan 2017. Good summary of word embeddings with interactive visualization tools, including word2viz word analogies explorer.
    9. FastText provides pre-trained word embeddings in many languages.

    Things to do:

    1. * Install Gensim in your environment (run "conda install gensim") and run the Gensim Word2vec tutorial.
    2. Experiment with Sense2vec with spaCy and Gensim, Source code a tool to compute word embeddings taking into account multi-word expressions and POS tags.
    3. Register to the Udacity Deep Learning course (free) by Vincent Vanhoucke, and study the chapter "Deep Models for Text and Sequences", then do Assignment 5 (ipynb notebook) "Train a Word2Vec skip-gram model over Text8 data".
    4. Continue with Assignment 6 (a ipynb notebook) "Train a LSTM character model over Text8 data".

  5. 11 Nov 2018: Parts of Speech Tagging (ipynb)

    Things to do:

    1. Read about the Universal Parts of Speech Tagset (About 2 hours)
    2. Install NLTK: if you have installed Anaconda, it is already installed. Make sure to download the corpora included with nltk.

      Q: How do you find out where your package is installed after you use easy_install?

      A: in the Python shell, type: import nltk; then type: nltk. You will get an answer like:

      >>> import nltk
      >>> nltk
      <module 'nltk' from 'C:\Anaconda\lib\site-packages\nltk\__init__.py'>
    3. Explore the Brown corpus of parts-of-speech tagged English text using NLTK's corpus reader and FreqDist object: Use the Universal tagset for all work (About 1 hour)
      • What are the 10 most common words in the corpus?
      • What are the 5 most common tags in the corpus?
    4. Read Chapter 5 of the NLTK book (About 3 hours)
    5. Advanced topics in POS tagging: we will get back to the task of POS tagging with different methods in the following chapters, for more advanced sequenced labeling methods (HMM), Deep Learning based methods using Recurrent Neural Networks, feature-based classifier methods for tagging (CRF), and as a test case for unsupervised EM techniques and Bayesian techniques. You can look at the source code of the nltk.tag module for a feeling of how the tag.hmm, tag.crf and tag.tnt methods are implemented.

      The following papers give a good feeling of the current state of the art in POS tagging:

    6. Read A good POS tagger in 200 lines of Python, an Averaged Perceptron implementation with good features, fast, reaches 97% accuracy (by Matthew Honnibal, 2013).
    7. Execute pos-tagging-skl.py, which implements a POS tagger using the Scikit-Learn model, with similar good features, fast, reaches 97% accuracy (80 lines)

  6. 18 Nov 2018: Basic Statistical Concepts for Supervised Machine Learning in NLP

    Things to do:

    1. * Bayesian concept learning from Tenenbaum 1999 - reported in Murphy 2012 Chapter 3. (I also warmly recommend this slightly related ICML 2018 lecture of Josh Tenenbaum - 1h11 Building Machines that Learn & Think Like People).
    2. * Read Deep Learning by Goodfellow, Bengio and Courville, 2016 Chapters 3 and 5. (About 5 hours)
    3. * Watch the 15mn video (ML 7.1) Bayesian inference - A simple example by Mathematical Monk.
    4. Make sure you have installed numpy and scipy in your Python environment. Easiest way is to use the Anaconda distribution.
    5. Read Introduction to statistical data analysis in Python - frequentist and Bayesian methods from Cyril Rossant, and execute the associated Jupyter notebooks. (About 4 hours)
    6. Learn how to use Scipy and Numpy - Chapter 1 in Scipy Lectures (About 5 hours)
    7. Write Python code using numpy, scipy and matplotlib.pyplot to draw the graphs of the Beta distribution that appear in the lecture notes (About 1 hour)
    8. Given a dataset for a Bernouilli distribution (that is, a list of N bits), generate a sequence of N graphs illustrating the sequential update process, starting from a uniform prior until the Nth posterior distribution. Each graph indicates the distribution over μ, the parameter of the Bernouilli distribution (which takes value in the [0..1] range). (About 2 hours)
    9. Learn how to draw Dirichlet samples using numpy.random.mtrand.dirichlet. A sample from a Dirichlet distribution is a multinomial distribution. Understand the example from the Wikipedia article on Dirichlet distributions about string cutting:
         import numpy as np
         import matplotlib.pyplot as plt
         s = np.random.dirichlet((10, 5, 3), 20).transpose()
         plt.barh(range(20), s[0])
         plt.barh(range(20), s[1], left=s[0], color='g')
         plt.barh(range(20), s[2], left=s[0]+s[1], color='r')
         plt.title("Lengths of Strings")
      (About 2 hours)
    10. Compute the MLE estimator μMLE of a binomial distribution Bin(m|N, μ).
    11. Mixture Priors: assume we contemplate two possible modes for the value of our Beta-Binomial model parameter μ. A flexible method to encode this belief is to consider that our prior over the value of μ has the form:
         μ ~ k1Beta(a, b) + k2Beta(c, d)
         where k1 + k2 = 1
         m ~ Bin(μ N)
      A prior over μ of this form is called a mixture prior - as it is a linear combination of simple priors.
      1. Prove that the mixture prior is a proper probabilistic distribution.
      2. Compute the posterior density over μ for a dataset where (N = 10, m=8, N-m=2) where k1=0.8 and k2=0.2 and the prior distributions are Beta(1,10) and Beta(10,1). Write Python code to draw the prior density of μ and its posterior density. (About 2 hours)
      3. Experiment with a very simple form of Stochastic Gradient Descent (SGD) with a custom loss function by running this notebook. (Notebook source here). More examples available on the Autograd project homepage.

  7. 25 Nov 18 Classification

    1. * Read Chapter 6: Learning to Classify Text of the NLTK Book (About 3 hours).
    2. * Read Generative and Discriminative Classifiers: Naive Bayes and Logistic Regression, Tom Mitchell, 2015. (About 3 hours)
    3. Watch (ML 8.1) Naive Bayes Classification a 15mn video on Naive Bayes Classification by Mathematical Monk and the following chapter (ML 8.3) about Bayesian Naive Bayes (20 minutes).
    Practical work:
    1. Explore the documentation of the nltk.classify module.
    2. Read the code of the NLTK Naive Bayes classifier and run nltk.classify.naivebayes.demo()
    3. Read the code of the NLTK classifier demos: names_demo and wsd_demo.
    4. Read the documentation on feature extraction in Scikit-learn.
    5. Run the example on document classification in Scikit-learn: Notebook (ipynb source).
    6. Experiment with the example of classifications in this iPython notebook (code) which shows how to run NLTK classifiers in a variety of ways.
    7. Experiment with the Reuters Dataset notebook (code) illustrating document classification with bag of words features and TF-IDF transformation.

  8. 02 Dec 18 Sequence Classification

    1. * Read Michael Collin's nodes on Tagging Problems, and Hidden Markov Models: POS tagging and Named Entity Recognition as tagging problems (with BIO tag encoding), generative and noisy channel models, generative tagging models, trigram HMM, conditional independence assumptions in HMMs, estimating the parameters of an HMM, decoding HMMs with the Viterbi algorithm.
    2. * Sequence Models and LSTM - PyTorch Tutorial.
    3. Introduction to RNNs (slides from Graham Neubig, Fall 2017).

    Things to do:

    1. Explain why the problem of decoding (see 2.5.4 in Tagging Problems, and Hidden Markov Models) requires a dynamic programming algorithm (Viterbi) while we did not need such a decoding step when we discussed classification using Logistic Regression and Naive Bayes?
    2. Implement Algorithm 2.5 (Viterbi with backpointers) from Tagging Problems, and Hidden Markov Models in Python. Test it on the Brown POS tagging dataset using MLE for tag transitions estimation (parameters q) and a discounting language model for each tag in the Universal taget for parameters e(x|tag) for each tag.
    3. Do the Assignment 3 from Richard Johannson's course on Machine Learning for NLP, 2014. Read the material assignment 3 material and Lecture 6: predicting structured objects.

      Start from the excellent Python implementation of the structured perceptron algorithm.

  9. 23 Dec 18 Syntax and Parsing
    1. * Context Free Grammars Parsing
    2. * Probabilistic Context Free Grammars Parsing
    3. * Michael Collins's lecture on CFGs and CKY
    4. * Michael Collins's lecture on Lexicalized PCFGs:
      1. Why CFGs are not adequate for describing treebanks: lack of sensitivity to lexical items + lack of sensitivity to structural preferences.
      2. How to lexicalize CFGs with Head propagation.
      3. How to parse a lexicalized PCFG.
    5. CKY parsing interactive demo in Javascript
    6. NLTK tools for PCFG parsing
    7. Notes on computing KL-divergence

  10. 30 Dec 18 Dependency Parsing
    1. * Dependency Parsing by Graham Neubig. Graham's teaching page with github page for exercises.
    2. * Dependency Parsing: Past, Present, and Future, Chen, Li and Zhang, 2014 (Coling 2014 tutorial)
    3. * NLTK Dependency Parsing Howto
    4. Parsing English with 500 lines of Python, an implementation by Matthew Honnibal of Training Deterministic Parsers with Non-Deterministic Oracles, Yoav Goldberg and Joakim Nivre, TACL 2013. (Complete Python code)
    5. Neural Network Dependency Parser, Chen and Manning 2014. A Java implementation of a Neural Network Dependency Parser with Unlabelled accuracy of 92% and Labelled accuracy of 90%.


Last modified 27 Jan 2019 Michael Elhadad