Example · 3 days
Applied NLP and the maths behind it
Classical NLP through transformers in Python, from TF-IDF to Hugging Face, with the linear algebra, probability and statistics that explain why it works.
For data scientists, analysts and ML engineers.
Day 1
Maths: linear algebra for embeddings and attention
Quarter day · NLP and the maths behind models · Intermediate
- Vectors, dot products, norms and cosine similarity.
- Matrices as transformations; matrix multiplication by hand and in NumPy.
- Projections and dimensionality reduction: PCA and what it keeps.
- Attention written as matrix operations.
Maths: probability and information theory
Quarter day · NLP and the maths behind models · Intermediate
- Distributions, conditional probability and Bayes’ rule.
- Softmax, temperature and sampling.
- Entropy, cross-entropy, KL divergence and perplexity.
- Exercise: compute a model’s loss on a sentence by hand.
NLP: Python for text and preprocessing
Quarter day · NLP and the maths behind models · Foundation
- Strings, Unicode, regular expressions and reading text at scale.
- Tokenisation, normalisation, stemming and lemmatisation.
- Stop words, and when removing them hurts.
- Lab: clean and tokenise a real corpus.
NLP: features, bag of words and TF-IDF
Quarter day · NLP and the maths behind models · Foundation
- Bag of words, n-grams and vocabularies.
- TF-IDF and what it rewards.
- Sparse matrices and similarity between documents.
- Lab: find similar documents with TF-IDF.
Day 2
NLP: text classification
Quarter day · NLP and the maths behind models · Intermediate
- Naive Bayes, logistic regression and linear SVMs on text.
- Imbalanced classes, precision, recall and thresholds.
- Error analysis: reading the mistakes.
- Lab: build and evaluate a classifier for a real labelling task.
NLP: topic modelling
Quarter day · NLP and the maths behind models · Intermediate
- LDA and NMF.
- Embedding-based topics with BERTopic.
- Choosing the number of topics and judging coherence.
- Lab: find the themes in a set of customer comments.
Maths: gradients and optimisation
Quarter day · NLP and the maths behind models · Intermediate
- Derivatives, the chain rule and backpropagation.
- Gradient descent, momentum and Adam.
- Learning rates, overfitting and regularisation.
- Lab: train a tiny model with gradients you wrote yourself.
NLP: word and sentence embeddings
Quarter day · NLP and the maths behind models · Intermediate
- Word2vec and GloVe: meaning from context.
- Analogies, bias and the limits of static vectors.
- Sentence embeddings and semantic search.
- Lab: compare TF-IDF and embeddings on the same search task.
Day 3
NLP: linguistic structure with spaCy
Quarter day · NLP and the maths behind models · Intermediate
- Part-of-speech tagging and dependency parsing.
- Named entity recognition, and training a custom entity.
- Rule-based matching alongside models.
- Lab: extract structured records from free text.
NLP: transformers with Hugging Face
Quarter day · NLP and the maths behind models · Intermediate
- Pretrained encoders and the Hugging Face ecosystem.
- Fine-tuning a small transformer for classification or NER.
- When a small fine-tuned model beats an LLM, on cost and accuracy.
- Lab: fine-tune and compare against the classical baseline.
Statistics for evaluating AI systems
Quarter day · NLP and the maths behind models · Intermediate
- Sampling error and how big a test set needs to be.
- Confidence intervals and bootstrapping for accuracy and judge scores.
- Paired comparisons: is version B really better than version A?
- Exercise: decide whether a reported improvement is real.
NLP: capstone pipeline
Quarter day · NLP and the maths behind models · Intermediate
- Put the pieces together: clean, extract, classify and search.
- Choosing classical, transformer or LLM for each step.
- Packaging the pipeline and evaluating it end to end.
- Present the pipeline and its numbers to the group.