Bagging to BERT- A tour of applied NLP
summary
Compares concrete methods on the IMDb 50k sentiment dataset and shows what actually moves the needle: a scikit-learn Count/TF‑IDF + Logistic baseline hits ~0.87 F1, reducing to ten-topic NMF topic vectors loses accuracy but gains interpretability (termite plots), averaging spaCy GloVe embeddings gives ~0.844, a simple LSTM sequence model around ~0.82, and a HuggingFace’s Transformers (BERT) pipeline nudges performance to ~0.88 — marginal improvement for much more complexity; includes scikit‑learn pipelines, spaCy access to embeddings, notes bias in corpora and practical preprocessing choices.
AI-written summary · about
chapters (10)
- 0:00Applied NLP introduction
- 11:01Bag-of-words to matrices
- 25:58Sentiment analysis baseline
- 29:10TF-IDF weighting
- 39:31Topic modeling with NMF
- 51:37Word embeddings with GloVe
- 1:04:55LSTM models and context
- 1:14:24Transformers explained
- 1:20:02BERT sentiment analysis
- 1:25:07Q&A
related talks
01h 19mGenerating Customers Insights with Topic Modelling and HuggingFace SetFit Method
21mPersonalization of Programming Education: An NLP-based Bi-dimensional(…)
40mPanel discussion "NLP"
31mTaking AI/ML to the next level: Text!
41mBetter Machine Learning with Less Data
01h 38mNavigating the World of NLP and Text Analytics
33mBulk Labelling Techniques
35mHow does text become data?