Word Embeddings for fun and profit in Gensim
summary
Uses Gensim to load Google News Word2Vec (300‑dim) and trains on 2,000 movie plots (170k words); builds document features by averaging word vectors and with doc2vec, then feeds logistic regression. tf‑idf and character n‑grams give ~42–44% baseline, averaging vectors and doc2vec reach ~50–52% accuracy, so neural embeddings give only ~5% uplift. Shows vector arithmetic working for analogies, reports doc2vec training ~25s but inference is slower.
AI-written summary · about
chapters (20)
- 0:00Word embeddings and Gensim
- 1:10Word embeddings benefits
- 3:57One-hot vectors definition
- 5:45Distributional hypothesis
- 7:12Word2vec gradient descent
- 9:40Softmax probabilities
- 11:26Vector arithmetic analogies
- 12:51Unsupervised learning applications
- 15:24Movie plot genre classification demo
- 17:33Averaging word vectors baseline
- 19:21Doc2Vec document tags
- 21:50Doc2Vec training and prediction
- 23:38TF-IDF and n-grams baselines
- 25:01Model tuning comparison
- 26:25Debugging embeddings
- 27:47Word embeddings use cases
- 28:53Other embedding applications
- 31:14NLP pre-processing tips
- 32:24LDA2Vec interpretable embeddings
- 33:06Loading external embeddings
related talks
01h 03mWord Embedding For Fun and Profit
39mNext generation of word embeddings in Gensim
01h 01mWord Embeddings - An Alternative & Efficient Approach to Search for Docs
45mDeep Learning for Natural Language Processing
40mText similiarity with the next generation of word embeddings in Gensim
28m"Embeddings: How Computers Learned to Read"
42mDocument Vectors in the Wild
39mBeyond word2vec: GloVe, fastText, StarSpace