😕 Confusion 📊 Prediction measure

perplexity

9.3

A number showing how well a computer model predicts text, where lower numbers mean better prediction

noun
/pərˈplɛksɪti/

Specialized

: perplejidad

To evaluate the performance of different models, researchers frequently measure their perplexity on standard datasets.

And so the, the, the main thing for me here is, you know, why is perplexity raising for the fourth time this year?

Examples

The team analyzed the perplexity of the language model to improve its accuracy in generating text.
In information theory, a model with lower perplexity is considered more effective at predicting outcomes.
We release both the Twitter search and the regular perplexity search a week apart.
And so the, the, the main thing for me here is, you know, why is perplexity raising for the fourth time this year?

The Prof G Pod – Scott Galloway
And the reality is, is that Google itself with ChatGPT and perplexity and all these AI things.
And look how many people in this room are now using AI, ChatGPT, perplexity, other things to look up stuff, not Google.
This, at least, will halve our perplexity.

Academic text (1998)

How Predictive

accuracy

169 3.8

perplexity

Specialized

9.3

predictability

17 3.3

Surface Forms

perplexity singular

perplexities plural

Morphology

perplexity = perplex (semi-transparent) = perplex + ity

The technical information-theoretic sense extends the basic idea of uncertainty/confusion to a formal numeric measure, which requires domain knowledge beyond the simple lexical derivation.

Etymology

Perplexity has the same root idea — per- 'thoroughly' + plex 'to weave' — as a picture of things tangled or hard to follow. In computing, perplexity became a number that shows how 'confused' a model is by some text, and lower values mean the model is less 'surprised' and better at prediction.