Word
perplexity
Phoneticsper·plex·i·ty
US/pɝˈpɫɛksəti/
UK/pəplˈɛksɪti/
Definition
n. a measurement of how surprised or confused an AI model is when reading a sequence of words.
n. a standard evaluation metric for language models, computed as the exponentiated cross-entropy loss, quantifying predictive uncertainty across tokens.
Examples
“Lower perplexity scores show that the language model finds the text natural and easy to predict.”
“The research team tracked training perplexity on a held-out dataset of classic novels.”
“Evaluating cross-domain generalization via token perplexity reveals how sharply predictive performance degrades outside the distribution of curated corpora.”
Examples
simple
“Lower perplexity scores show that the language model finds the text natural and easy to predict.”
contextual
“The research team tracked training perplexity on a held-out dataset of classic novels.”
complex
“Evaluating cross-domain generalization via token perplexity reveals how sharply predictive performance degrades outside the distribution of curated corpora.”
Real-World Examples
“Perplexity summarises how "surprised" a model is by a token stream: lower perplexity means the model assigned high probability to the tokens that actually occurred.” “A model with low perplexity is "less surprised" by real text, meaning it anticipated the words that appeared.” Etymology
From Latin perplexitas ('entanglement, confusion'), from perplexus ('confused, intricate'). Introduced as an information-theoretic language evaluation metric by Frederick Jelinek at IBM in the 1970s.
Etymology adapted from Wiktionary, available under CC BY-SA 4.0.
Domains
Scan code
englishreference.com/q/perplexity