The effect of learner errors on POS tag errors during automatic POS tagging

Abstract

The use of a corpus as a language resource is enhanced when it is part of speech (POS) tagged. There are a number of computer tagging programs (taggers) available. In order to decide which taggers to use to POS tag the Tswana Learner English Corpus (TLEC), the performance of three taggers (TOSCA-ICLE, Brill tagger, and CLAWS) was evaluated on a small sample of the corpus. An evaluation of an unedited version of the sample indicated that learner spelling errors contributed substantially to tagging errors. All spelling errors were then corrected and the edited sample was retagged. With the spelling errors removed, the performance of all three taggers improved. This article reports on the influence of the other learner errors on the performance of the taggers. It was found that 38% of the tag errors made by CLAWS on the edited sample were due to learner errors. In the case of TOSCA 19%, and in the case of Brill 14%, of the tag errors were due to learner errors.


(S/ern Af Linguistics & Applied Language Stud: 2002 20(4): 325-335)