[Corpora-List] Spellchecker evaluation corpus

A.P.J. van den Bosch Antal.vdnBosch at uvt.nl
Thu Apr 14 10:40:51 CEST 2011

Dear Stefan,

All good points, but when you say

> - several collections of misspelled words along with a defined context size of differing languages to evaluate spelling error detectors and correctors

what do you mean with a defined context size? What seems to be missing from your list is what I think should be the ultimate evaluation setting: _full_ texts with _all_ errors annotated.

Error list evaluations cannot measure the false alarm rate or precision of your spelling error detector: how often does it think it has found an error which isn't one? Put in another way, an algorithm with a great recall/accuracy on an error list may actually be an over-enthousiastic system that flags many normal words as errors as well.

For fully-automatic correction and corpus cleanup this is quite vital - does your method do more harm than good? But also interactive spellcheckers could do with a higher precision; as one of the most widely used pieces of language technology worldwide, it's not particularly loved for its low precision.


-- Antal van den Bosch Antal.vdnBosch at uvt.nl http://ilk.uvt.nl/~antalb/ ILK / Tilburg center for Cognition and Communication, Tilburg University

More information about the Corpora mailing list