A course on the modern aspects of Natural Language Processing. You will learn how today's NLP systems work end to end — from tokenization and language modelling to transformers and large language models, and how they are fine-tuned, retrieved from, aligned, served, and evaluated — and how to build, probe, and critically assess them.
Lectures are on Thursdays. Teaching team: J. Yang (responsible), A. Anand. Assessment: 20% individual assignments + 80% group project.
| Date | Morning | Afternoon | Assignments |
|---|---|---|---|
| 12 Nov | Intro to NLP + LogisticsYangCEG Lecture Hall E | — | — |
| 19 Nov | Lexical Analysis + Language ModellingAnandCEG Lecture Hall E | — | A1 |
| 26 Nov | Text Classification + EmbeddingsAnandCEG Lecture Hall E | Syntax and SemanticsYangEEMCS Chip | A1 due |
| 3 Dec | Neural LM & Contextual EmbeddingsYangEEMCS Chip | Discriminative TaskAnandCEG Lecture Hall E | A2 |
| 10 Dec | Generative ModellingAnandCEG Lecture Hall E | Text Generation TasksYangEEMCS Chip | A2 due |
| 17 Dec | Retrieval-based LMsYangCEG Lecture Hall E | AlignmentguestEEMCS Chip | A3 |
| 7 Jan | InterpretabilityAnandCEG Lecture Hall E | AgentsguestEEMCS Chip | A3 due |
| 14 Jan | LLM ServingAnandCEG Lecture Hall E | RobustnessYangEEMCS Chip | final project due |
BPE, a bigram (add-k) LM, perplexity per-token vs per-byte, compared to a TULIP-served LLM (Gemma 4).
LLM summarisation + an LLM-as-judge benchmark + human-eval alignment.
Generate → score → attribution → lexical/syntactic analysis.
Membership-inference & extraction attacks and defences on a small fine-tuned LM (Enron).
How context quality shapes RAG on multi-hop QA; hard-negative dense retrieval with ADORE (2WikiMultiHopQA).
Steer a summarizer's moral framing at inference with contrastive activation addition (CAA) — no fine-tuning (EMONA / Moral Foundations).
Browse the briefs above for the kind of scope we expect, then propose your own NLP project — send a short proposal to the teaching team by 10 December.
Download the course starter kit — loaders, a submission checker, and a data sample (A1 bundles a WikiText-103 sample, so there is nothing extra to fetch). Per-assignment kits are also on Brightspace.
We are running a TULIP pilot for the course: large models served OpenAI-compatibly over eduVPN with your personal key. Docs at docs.tulip.tudelft.nl.
Content note: some datasets (e.g. toxicity comments, Enron e-mails) contain offensive or sensitive material — handle with care.