BPE
Tokenisation that builds a vocabulary by merging frequent character pairs.
Overview
Byte Pair Encoding starts from individual characters and repeatedly merges the most frequent adjacent pair, so common words become one token and rare ones are assembled from parts. It solves the vocabulary problem: a word-level vocabulary cannot cover every name and inflection, and a character-level one makes sequences unnecessarily long. It exists as the scheme most language models still use, with its vocabulary size chosen rather than learned.
Technology Graph
How BPE connects to the rest of the stack.
Applied to
- Tokenization technique Techniques BPE used for Tokenization High ✓ source
Tools
Software that implements or integrates with BPE.