NLP: Start Here
What natural language processing covers today, the tasks and tools a junior NLP engineer uses, and how this track connects to the LLM topic and its game season.
NLP in the age of large language models
Natural language processing (NLP) builds systems that read, understand and generate human language. Large language models (LLMs) now solve many NLP tasks with a prompt, but that has made the fundamentals more important, not less: someone has to decide when a small fine-tuned model is better (cheaper, faster, private, more accurate on a narrow task), how text is tokenised, how outputs are evaluated, and how models are adapted and aligned.
| Task | Example | Typical approach today |
|---|---|---|
| Text classification | route a patient message: billing, prescription, urgent… | TF-IDF baseline → fine-tuned small encoder (or an LLM with few-shot prompts) |
| Sequence labelling (NER) | find names, medicines and dates to redact or extract | fine-tuned encoder with BIO tags |
| Semantic search | find policy paragraphs matching a question | sentence embeddings + vector index |
| Generation | summarise a consultation, draft a reply | LLM (prompted, RAG, or fine-tuned), carefully evaluated |
| Translation, speech | interpreter support, voice notes | multilingual / speech models |
Prerequisites: the Transformers and Transfer learning topics, and the core course's LLM topic (prompting, RAG, evaluation), which this track complements.
Tools, and how this track is organised
Standard tools: Hugging Face transformers, datasets, tokenizers and evaluate; sentence-transformers for embeddings; spaCy for fast pipelines (tokenising, NER); scikit-learn for baselines; PEFT/TRL for LoRA, instruction tuning and preference training; vLLM or llama.cpp for serving LLMs.
| Topics | Game |
|---|---|
| Text & tokenisation · Embeddings · Classification & sequence labelling | N1: The patient-message inbox |
| Language models, decoding & evaluation · Training and aligning LLMs | N2: Summaries clinicians can trust |
Hands-on: Colab notebooks for this track
11: fine-tune DistilBERT to classify messages and to tag entities (NER), against a TF-IDF baseline. 12: decoding strategies, perplexity and ROUGE, then LoRA instruction-tuning of a small language model.
Interactive: this part runs in your browser.
Quiz
1. Why still learn NLP fundamentals when LLMs can do many tasks with a prompt?
2. Finding every medicine name in a message is…
3. Which library provides LoRA and preference-training utilities?
📚 Further reading
Trusted, mostly free sources: official documentation, well-known textbooks and courses. When you want to confirm or go deeper on anything here, start with these.
- Speech and Language Processing (Jurafsky & Martin, 3rd ed. draft) ↗ The standard NLP textbook, free online
- Hugging Face LLM / NLP Course ↗ Free course on transformers, tokenizers and fine-tuning
- CS224N: NLP with Deep Learning (Stanford) ↗ Lectures and notes