An NLP course that lets you swing before explaining the rules

What it does This repo holds the Jupyter notebooks for a 2019 University of San Francisco Data Science masters course. It walks through classical NLP—TF-IDF, Naive Bayes, and n-grams—up to modern deep learning including RNNs, Transformers, and multi-head attention, using sklearn, nltk, pytorch, and fastai. The syllabus also dedicates a full module to bias, ethics, and fake information in the attention economy.

The interesting bit The pedagogy is deliberately inverted. You start running code in “black boxes” and only later peel back the layers to understand SVD, attention masking, or label smoothing. The instructors use a baseball analogy: kids play the game before memorizing the rulebook, so you focus on what techniques do before learning what they are.

Key highlights

• Full arc from sparse-matrix topic modeling (NMF, SVD) to Transformer-based translation
• Dedicated module on bias in word embeddings, types of bias, and the attention economy
• Originally taught live; lecture videos and a companion blog post are linked
• Mixes classical libraries (sklearn, nltk) with modern deep-learning stacks (pytorch, fastai)
• Sentiment classification and translation are the primary hands-on tasks

Caveats

• The curriculum is fixed at its 2019 syllabus, so the deep-learning examples stop at the Transformer architecture
• Early notebooks are intentionally opaque; the README warns that you are “not supposed to” understand everything at first

Verdict Best for Python developers who want a structured, code-first path from sparse matrices to Transformers without prerequisite theory. Skip it if you need the underlying mechanics explained before you run the code.