CS 533: Natural Language Processing (Spring 2020)

Coronavirus update. All classes after Spring Recess (i.e., from March 25) will be live-streamed within the usual time slot until further notice. The project presentations will also be live-streamed if in-person delivery is infeasible due to the coronavirus situation. Please see Canvas for details on how to join online classes.

Instructor: Karl Stratos
TA: Zuohui Fu (office hours: Tuesday 3:30-4:30pm, Hill 273)
Time and location: Wednesday 12-3pm at BE 252
Instructor office hours: Wednesday 3:20-4:30pm at Tillett 111H

Course description. This project-centered graduate course will cover technical foundations of modern NLP. Students are expected to start working on course projects immediately from the beginning of the course and throughout, culminating in (1) in-class project presentations and (2) written reports that aspire to conference publication level. The course will have two parts that happen in parallel. The first part is standard lecture-based classes in which the instructor exposes students to fundamental concepts and applications in the field. The second part is continual discussions and brainstorming about course projects and self-initiated research efforts. There is no required textbook: all materials are publicly available online resources.

Please use the Canvas site to ask questions regarding lectures/homeworks/projects, to submit assignments, and to find announcements.

  1. Achieving an understanding of the foundational concepts and tools used in modern NLP
  2. Obtaining an ability to critically read and accurately evaluate conference papers in NLP
  3. Finding new research projects that persist beyond this course

Audience and prerequisites. No previous exposure to NLP is assumed. However, this is a fast-paced course designed for self-motivated graduate or advanced undergraduate students with a solid technical background in probability and statistics, calculus, and linear algebra. Technical requirements include:
  1. Probabilistic reasoning (e.g., What is the conditional probability of Y=y given X=x, assuming the knowledge of a joint distribution over X and Y?)
  2. Intimate and intuitive understanding of matrix and vector operations (e.g., What is the shape of a matrix product? How similar are two vectors?)
  3. Mathematical notions in optimization (e.g., What does it mean for a function to have zero derivative at a certain point?)
If you cannot complete A1 comfortably, you may need to consult with the instructor about whether your background meets the prerequisites. Significant programming experience in Python is necessary for programming assignments and course projects.

  1. Assignments: 50% (10% per assignment)
  2. Project: 40% (written report 30%, presentation 10%)
  3. Participation: 10%
The assignment report must be written in LaTeX using the provided assignment report template. Similarly, the project report must be written in LaTeX using the provided project report template and will be reviewed by the instructor like a conference submission.

Project timeline.
  1. Proposal (due 3/24) : submit an initial proposal using this template.
  2. Milestone (due 4/15): submit an informal 1-2 page progress report.
  3. Presentation (tentatively 4/29): in-class presentation
  4. Final report (due 5/4): submit a final report

Tentative plan.
Date Topics Readings Assignments
Week 1 (January 22) Logistics, Introduction, Language Modeling Michael Collins' notes on n-gram models and log-linear models A1 [code] (Due 2/4)
Week 2 (January 29) Deep Learning for NLP: Neural Language Modeling Colah's blogs on deep learning and LSTMs, NLM papers using feedforward (Bengio et al., 2003), recurrent (Mikolov et al., 2010; Melis et al., 2018), and attention-based (GPT-2) architectures
Week 3 (February 5) Deep Learning for NLP: Conditional Neural Language Modeling BLEU, input-feeding attention, Google's NMT, summarization, copy mechanism, data-to-text generation A2 [code] (Due 2/18)
Week 4 (February 12) Deep Learning for NLP: Backpropagation, Self-Attention, Representation Learning by Language Modeling Backpropagation, Transformer (note), ELMo, BERT
Week 5 (February 19) Structured Prediction in NLP: Tagging Michael Collins' notes on HMMs, CRFs, and forward-backward, neural architectures for sequence labeling (Collobert et al., 2011; Lample et al., 2016) A3 [code] (Due 3/10)
Week 6 (February 26) Structured Prediction in NLP: Constituency and Dependency Parsing constituency parsing (Michael Collins' notes on PCFGs and inside-outside algorithm; Kitaev and Klein, 2018), transition-based dependency parsing (Nivre, 2008; Chen and Manning, 2014) graph-based dependency parsing (Eisner, 1996; Kiperwasser and Goldberg, 2016)
Week 7 (March 4) Unsupervised Learning in NLP: Latent-Variable Generative Models and the EM Algorithm David McAllester's notes on EM, EM for Naive Bayes model (MRS, 2019; Michael Collins' notes), EM for PCFGs (Lari and Young, 1990)
Week 8 (March 11) Unsupervised Learning in NLP: Autoencoders and VAEs Section 1 and Appendix A of this note, useful facts about latent-variable generative models, VAEs for NLP (Bowman et al., 2016; Pelsmaeker and Aziz, 2019) A4 [code] (Due 3/31)
Spring Recess
Week 9 (March 25) Information Extraction in NLP Document weighting schemes (MRS Chapter 6, TFIDF, BM25), entity linking (Ling et al., 2015; Logeswaran et al., 2019; Gillick et al., 2019; Kolitsas et al., 2018), retrieval-based question answering (Chen et al., 2017; Lee et al., 2019), coreference resolution (Lee et al., 2017), relation extraction (slides by Huck and Fraser)
Week 10 (April 1) Special Topics: Large-Scale Transfer Learning for NLP Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer Proposal due 3/31, A5 (Due 4/21)
Week 11 (April 8) Special Topics: Data Annotation and Question Answering BREAK It Down: A Question Understanding Benchmark, HOTPOTQA: A Dataset for Diverse, Explainable Multi-hop Question Answering
Week 12 (April 15) Milestone Presentations Milestone due 4/14
Week 13 (April 22) Special Topics: Parallel Decoding, Discriminative vs Generative Models Mask-Predict: Parallel Decoding of Conditional Masked Language Models, Your Classifier is Secretly an Energy Based Model and You Should Treat it Like One
Week 14 (April 29) Project Presentations

Other resources.
  1. Speech and Language Processing (3rd edition) by Dan Jurafsky and James H. Martin
  2. A Primer on Neural Network Models for Natural Language Processing by Yoav Goldberg
  3. Natural Language Processing by Jacob Eisenstein