Lectures are on Thursdays. Teaching team: J. Yang (responsible), C. Lofi, A. Anand. Assessment: 20% individual assignments + 80% group project.
| Date | Morning | Afternoon | Assignment out |
|---|---|---|---|
| 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 | A2 |
| 3 Dec | Discriminative TaskAnandCEG Lecture Hall E | Neural LM & Contextual EmbeddingsYangEEMCS Chip | A3 |
| 10 Dec | Generative ModellingAnandCEG Lecture Hall E | Text Generation TasksYangEEMCS Chip | A4 |
| 17 Dec | Retrieval-based LMsYangCEG Lecture Hall E | AlignmentguestEEMCS Chip | A5 |
| 7 Jan | InterpretabilityAnandCEG Lecture Hall E | AgentsguestEEMCS Chip | — |
| 14 Jan | LLM ServingAnandCEG Lecture Hall E | RobustnessYangEEMCS Chip | final project due |
BPE, n-gram LM, perplexity per-token vs per-byte, GPT-2.
Precision/recall/F1, per-subgroup fairness, error taxonomy.
Bag-of-words → embeddings → DistilBERT.
Fine-tune GPT-2; canaries & a memorisation/extraction attack.
Batched client, latency sweep, token/compute budget.
Per-layer probes / attribution. Findings-only.
LLM summarisation + an LLM-as-judge benchmark + human-eval alignment.
Context-quality ablations in retrieval-augmented generation.
Generate → score → attribution → lexical/syntactic analysis.
Membership-inference / extraction attacks and defences.
Closed-book vs RAG; temporal reasoning over lists.
Improve NLI veracity prediction + claim decomposition.
Semantic-parsing + IR-based KBQA with pretrained models.
Join the competition to get the data. Binary toxicity + identity subgroups; comment text for LM work.
Kaggle data loader, submission checker, requirements.
Join the Kaggle competition, set up API credentials, then the starter loader derives the splits:
from load_data import load_textclf, load_lm # starter.zip data = load_textclf() # train / validation / test (toxicity + identity subgroups) train, heldout = load_lm() # raw comment text for language-model work # models come from HuggingFace as usual (transformers)
Content note: the comments can contain offensive language.