Skip to main content

Natural language processing

From raw customer reviews to a document classifier good enough to deploy, in ten focused modules.

Course duration: 7h​

What you will learn​

  • Clean and normalise real text without destroying what carries meaning
  • Tokenise into words, subwords, BPE and SentencePiece, and predict the cost of each choice
  • Build a TF-IDF baseline and know when it is enough to ship
  • Read the geometry of word embeddings and spot the biases they encode
  • Use contextual encoders (BERT and Sentence-BERT) for classification and semantic search
  • Fine-tune a pretrained model on your own corpus with the Hugging Face Transformers library
  • Extract named entities from noisy reviews and evaluate them per entity type
  • Summarise and answer questions on a passage, and see through the ROUGE score
  • Compare three approaches on the same task: TF-IDF, embeddings and fine-tuning
  • Pick a model on the Hub with an eye on size, licence and language coverage

Prerequisites​

  • Python (course 02)
  • Supervised learning (course 04)
  • Transformers architecture (course 12) recommended

Course modules​

  1. Text processing pipeline: cleaning and normalization
  2. Tokenization: words, subwords, BPE, SentencePiece
  3. Bag of words, TF-IDF and their limits
  4. Word2Vec, GloVe and the geometry of meaning
  5. Contextual embeddings and pretrained models
  6. Text classification and sentiment analysis
  7. Named entity recognition
  8. Summarization and question answering
  9. The Transformers library in practice
  10. Project: a document classifier in the course language

The corpus that runs through the course​

Every module works on the same corpus: about 10,000 English customer reviews of consumer products, each labelled with a 1-to-5 star rating. You can use a public dataset in that shape (for instance the Amazon Fine Food Reviews or the Yelp Open Dataset filtered to English) or describe an equivalent internal corpus. The choice is deliberate: reviews are short but noisy, they contain contractions, misspellings, emojis and product names, and the star rating gives a supervised signal for every downstream task in the course.

The pipeline is built up one module at a time. Module 1 cleans the text and normalises Unicode. Module 2 turns it into tokens. Module 3 fits a TF-IDF plus logistic regression baseline you will keep referring back to. Modules 4 and 5 add dense representations, first static then contextual. Module 6 fine-tunes an encoder for sentiment. Module 7 pulls product and place mentions out of the same reviews. Module 8 tries to summarise them and to answer questions on a paragraph. Module 9 assembles everything with the Transformers library. Module 10 closes the loop with a full project that compares the three approaches on cost, latency and quality — and pays attention to the specifics of English (contractions, capitalisation, negation scope).

Assessment and certificate​

The course ends with a 40-question exam covering every module. On success, a certificate of completion is issued; its number can be verified by anyone on the platform.

Free courses, by contrast, end with a 5-question quiz and a certificate preview, without certification.