DecodeAI
NLP Fundamentals
Free

Tokenization & Text Representations

From words to subwords — tradeoffs that affect every NLP system.

Tokenization & Text Representations

Tokenization decides the atomic units your model sees.

Options

  • Whitespace / word tokens: simple, brittle on morphology
  • Subwords (BPE, WordPiece, Unigram): handle rare words
  • Bytes / characters: flexible, longer sequences

Embeddings

Map tokens to dense vectors. Static embeddings (word2vec/GloVe) vs contextual (transformers).