CS-552: Modern Natural Language Processing

Course Description

Natural language processing is ubiquitous in modern intelligent technologies, serving as a foundation for language translators, virtual assistants, search engines, and many more. In this course, we cover the foundations of modern methods for natural language processing, such as word embeddings, recurrent neural networks, transformers, and pretraining, and how they can be applied to important tasks in the field, such as machine translation and text classification. We also cover issues with these state-of-the-art approaches (such as robustness, interpretability, sensitivity), identify their failure modes in different NLP applications, and discuss analysis and mitigation techniques for these issues.

Quick access links:

• Platforms
• Lecture Schedule
• Exercise Schedule
• Grading
• Contact

Class

Platform	Where & when
Lectures	Wednesdays: 11:15-13:00 [STCC - Cloud C] & Thursdays: 13:15-14:00 [CE16]
Exercises Session	Thursdays: 14:15-16:00 [CE11]
Project Assistance (not every week)	Wednesdays: 13:15-14:00 [STCC - Cloud C]
QA Forum & Annoucements	Ed Forum [link]
Grades	Moodle [link]

All lectures will be given in person and live streamed on Zoom. The link to the Zoom is available on the Ed Forum (pinned post). Beware that, in the event of a technical failure during the lecture, continuing to accompany the lecture live via zoom might not be possible.

Recording of the lectures will be made available on SwitchTube. We will reuse some of last year's recordings and we may record a few new lectures in case of different lecture contents.

Lecture Schedule

Week	Date	Topic	Suggested Reading	Instructor
Week 1	21 Feb 22 Feb	Introduction | Building a simple neural classifier [slides] Neural LMs: word embeddings [slides]	Suggested reading: Introduction to natural language processing, chapter 3.1 - 3.3 & chapter 14.5 - 14.6 Efficient Estimation of Word Representations in Vector Space GloVe: Global Vectors for Word Representation Enriching word vectors with subword information Advances in pre-training distributed word representations	Antoine Bosselut
Week 2	Feb 29 Feb	LM basics | Neural LMs: Fixed Context Models [slides] Neural LMs: RNNs, Backpropagation, Vanishing Gradients; LSTMs [slides]	Suggested reading: Introduction to natural language processing, chapter 6.1-6.4 A Neural Probabilistic Language Model On the difficulty of training recurrent neural networks Introduction to natural language processing, chapter 3.1 - 3.3 & chapter 18.3, 18.4	Antoine Bosselut
Week 3	6 Mar 7 Mar	Seq2seq + decoding + attention | Transformers [slides] Transformers + Greedy Decoding; GPT [slides]	Suggested reading: Attention Is All You Need The Annotated Transformer The illustrated transformer GPT: Improving language understanding by generative pre-training GPT2: Language Models are Unsupervised Multitask Learners	Antoine Bosselut
Week 4	13 Mar 14 Mar	[Online only]Pretraining: ELMo, BERT, MLM, task generality | Transfer Learning: Introduction [slides] Assignment 1 Q&A	Suggested reading: Elmo: Deep Contextualized Word Representations BERT: BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding RoBERTa: RoBERTa: A Robustly Optimized BERT Pretraining Approach ELECTRA: ELECTRA: Pre-training Text Encoders as Discriminators Rather Than Generators Transfer Learning in Natural Language Processing T5: Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer BART: BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension	Antoine Bosselut Simin Fan
Week 5	20 Mar 21 Mar	Transfer Learning: Dataset Biases [slides] Generation: Task [slides]	Suggested reading: -	Antoine Bosselut
Week 6	27 Mar 28 Mar	Generation: Decoding and Training [slides] Generation: Evaluation [slides]	Suggested reading: Decoding: On Using Monolingual Corpora in Neural Machine Translation Decoding: Google's Neural Machine Translation System: Bridging the Gap between Human and Machine Translation Decoding: Improving LSTM-based Video Description with Linguistic Knowledge Mined from Text Decoding: A Diversity-Promoting Objective Function for Neural Conversation Models Decoding: A Deep Reinforced Model for Abstractive Summarization Decoding: Deep Communicating Agents for Abstractive Summarization Decoding: Learning to Write with Cooperative Discriminators Decoding: Hierarchical Neural Story Generation Decoding: Discourse Understanding and Factual Consistency in Abstractive Summarization Decoding: Plug and Play Language Models: A Simple Approach to Controlled Text Generation Decoding: The Curious Case of Neural Text Degeneration Decoding: Generalization through Memorization: Nearest Neighbor Language Models Training: Scheduled Sampling for Sequence Prediction with Recurrent Neural Networks Training: Sequence Level Training with Recurrent Neural Networks Training: Google's Neural Machine Translation System: Bridging the Gap between Human and Machine Translation Training: Deep Reinforcement Learning-based Image Captioning with Embedding Reward Training: Self-critical Sequence Training for Image Captioning Training: Improved Image Captioning via Policy Gradient Optimization of SPIDEr Training: Sentence Simplification with Deep Reinforcement Learning Training: A Deep Reinforced Model for Abstractive Summarization Training: Deep Communicating Agents for Abstractive Summarization Training: Discourse-Aware Neural Rewards for Coherent Text Generation Training: Polite Dialogue Generation Without Parallel Data Training: Paraphrase Generation with Deep Reinforcement Learning Training: The Curious Case of Neural Text Degeneration	Antoine Bosselut
		*EASTER BREAK*
Week 7	10 Apr 11 Apr	In-context Learning - GPT-3 + Prompts | Instruction Tuning [slides] Project Description	Suggested reading: -	Antoine Bosselut
Week 8	17 Apr 18 Apr	[Online only]Scaling laws | Model Compression [slides] No class	Suggested reading: Scaling laws for neural language models Training compute-optimal large language models	Antoine Bosselut
Week 9	24 Apr 25 Apr	Ethics in NLP: Bias / Fairness and Toxicity, Privacy, Disinformation [slides] No class (Project work; A1 Grade Review Session)	Suggested reading: Ethics in NLP United Nations recommendations/overview on responsible AI practice	Anna Sotnikova
Week 10	1 May 2 May	Tokenization: BPE, WP, Char-based | Multilingual LMs [slides] Guest Lecture: Kayo Yin	Suggested reading: Between words and characters: A brief history of open-vocabulary modeling and tokenization in NLP Byt5: Towards a token-free future with pre-trained byte-to-byte models Neural machine translation of rare words with subword units Unsupervised cross-lingual representation learning at scale Emerging cross-lingual structure in pretrained language models Mad-x: An adapter-based framework for multi-task cross-lingual transfer Composable sparse fine-tuning for cross-lingual transfer The State of Multilingual AI	Negar Foroutan Kayo Yin
Week 11	8 May 9 May	Syntactic and Semantic Tasks (NER) | Interpretability: BERTology No class (Project work; A2 Grade Review Session)	Suggested reading: -	Gail Weiss
Week 12	15 May 16 May	Reading Comprehension | Retrieval-augmented LMs No class (Project work; A2 Grade Review Session)	Suggested reading: Squad: 100,000+ questions for machine comprehension of text Natural questions: a benchmark for question answering research Dense passage retrieval for open-domain question answering Retrieval augmented language model pre-training Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks Toolformer: Language models can teach themselves to use tools React: Synergizing reasoning and acting in language models Improving language models by retrieving from trillions of tokens Augmented language models: a survey	Antoine Bosselut
Week 13	22 May 23 May	Multimodality: L & V Looking forward	Suggested reading: -	Syrielle Montariol Antoine Bosselut
Week 14	29 May 30 May	No class (Project work; A3 Grade Review Session)

Exercise Schedule

Week	Date	Topic	Instructor
Week 1	22 Feb	Setup + Word embeddings [code]	Mete Ismayilzada
Week 2	29 Feb	Word embeddings review Language and Sequence-to-sequence models [code]	Mete Ismayilzada Badr AlKhamissi
Week 3	6 Mar	Assignment 1 Q&A	Mete Ismayilzada
Week 3	7 Mar	Language and Sequence-to-sequence models review Attention + Transformers [code]	Badr AlKhamissi
Week 4	13 Mar	[Online only] Pretraining S2S: BART, T5 [[slides][4s]]	Antoine Bosselut
Week 4	14 Mar	Attention + Transformers review Pretraining and Transfer Learning Pt. 1 [code]	Badr AlKhamissi Simin Fan
Week 5	20 Mar	No lecture	-
Week 5	21 Mar	Pretraining and Transfer Learning Pt. 1 review Transfer Learning Pt. 2 [code]	Simin Fan
Week 6	27 Mar	Assignment 2 Q&A	Simin Fan, Silin Gao
Week 6	28 Mar	Transfer Learning Pt. 2 review Text Generation & Assignment 2 Q&A [code]	Simin Fan Deniz Bayazit, Silin Gao
		EASTER BREAK
Week 7	10 Apr	Assignment 3 Q&A	Badr AlKhamissi Deniz Bayazit
Week 7	11 Apr	Text Generation review In-context Learning [code]	Deniz Bayazit Mete Ismayilzada
Week 8	17 Apr	No lecture	-
Week 8	18 Apr	Assignment 3 Q&A A1 Grade Review Session	Badr AlKhamissi Deniz Bayazit Mete Ismayilzada
Week 9	24 & 25 Apr	Project	TA meetings on-demand
Week 10	1 & 2 May	Project	TA meetings on-demand
Week 11	8 & 9 May	Project Milestone 1 Feedback	TA meetings on-demand
Week 12	15 & 16 May	Project	TA meetings on-demand
Week 13	22 May	Project	Badr AlKhamissi Deniz Bayazit
Week 13	23 May	A3 Grade Review Session	TA meetings on-demand
Week 14	30 May	Project Milestone 2 Feedback	TA meetings on-demand

Exercises Session format:

• TAs will provide a small discussion over the last week's exercises, answering any questions and explaining the solutions. (10-15mins)
• TAs will present this week's exercise. (5mins)
• Students will be solving this week's exercises and TAs will provide answers and clarification if needed.

Note: Please make sure you have already done the setup prerequisites to run the coding parts of the exercises. You can find the instructions here.

Grading:

Your grade in the course will be computed according to the following guidelines.

Submission Format

Assignment and project release annoucements will be on Ed. Your work will be submitted as a repository created by GitHub classroom. Clicking the assignment link (announced on its release date) will automatically create a repository under your username (ensure it matches the one on the CS-552 GitHub registration form). Your last push to the repository will be considered as your final submission, with its timestamp determining any late days (see below for the policy).

All large files such as model checkpoints need to be pushed to the repository with Git LFS. Large files can take time to upload, therefore please avoid last-minute uploads that can create potential submission delays. We also propose to use Colab as a free GPU resource. You can find tutorials on all of these resources here.

Late Days Policy

All assignments and milestones are due at 23:59 on their due date. As we understand that circumstances can make it challenging to abide by these due dates, you will receive 7 late days over the course of the semester to be allocated to the assignments and project milestones as you see fit. No further extensions will be granted. The only exception to this rule is for the final report, code, and data. No extensions will be granted beyond June 14th. We will automatically calculate the late days according to your last commit; hence you don't have to inform us. For group projects, when everyone has some late days, we will deduct individually from everyone. In the scenario where one person has no more late days, that student will lose points for the late submission. The other students in the team will continue to use their late days (i.e. no points will be deducted from them). After you have used all your allotted late days, the penalty policy is a 25% deduction of the grade per day.

Assignments (40%):

There will be three assignments throughout the course. They will be released and due according to the following schedule:

Assignment 1 (10%)

• Released: 28 February 2024
• Due: 17 March 2024
• Grade released: 14 April 2024
• Grade review sessions: 18 and 25 April 2024

Assignment 2 (15%)

• Released: 20 March 2024
• Due: 7 April 2024
• Grade released: 5 May 2024
• Grade review sessions: 9 and 16 May 2024

Assignment 3 (15%)

• Released: 3 April 2024
• Due: 21 April 2024
• Grade released: 19 May 2024
• Grade review sessions: 29 and 30 May 2024

Project (60%):

The project will be divided into 2 milestones and a final submission. Each milestone will be worth 15% of the final grade with the remaining 30% being allocated to the final report. Each team will be supervised by one of the course TAs or AEs.

More details on the content of the project and the deliverables of each milestone will be released at a later date.

Milestone 1:

• Due: 5 May 2024

Milestone 2:

• Due: 28 May 2024

Final Deliverable:

• The final report, code, and date will be due on June 14th. Students are welcome to turn in their materials ahead of time, as soon as the semester ends.

• Due: 14 June 2024

Contacts

Please email us at nlp-cs552-spring2024-ta-team [at] groupes [dot] epfl [dot] ch for any administrative questions, rather than emailing TAs individually. All course content questions need to be asked via Ed.

Lecturer: Antoine Bosselut

Teaching assistants: Negar Foroutan Eghlidi, Badr AlKhamissi, Deniz Bayazit, Beatriz Borges, Zeming (Eric) Chen, Simin Fan, Silin Gao, Mete Ismayilzada