Building a Large Annotated Corpus of English: The Penn Treebank

2.3 The POS tagging process

1. How was the tagged corpus produced?
   Automatic tagging (initially Church’s stochastic PARTS tagger, mapped to the Penn tagset; later a cascade of stochastic and rule-based taggers), then manual correction by annotators in a GNU Emacs interface.
   • Marcus et al., 1993, Building a Large Annotated Corpus of English (Computational Linguistics), §2.3
2. Why did annotators append corrected tags (word/OLD*/NEW) rather than replace them?
   To make recurring errors of the automatic stage easy to identify, giving confusion matrices useful for designing better taggers.
   • Marcus et al., 1993, Building a Large Annotated Corpus of English (Computational Linguistics), §2.3
3. Why did mapping PARTS output to the Penn tagset introduce errors?
   PARTS didn’t make some Penn distinctions: infinitives vs non-third-person present verbs (VB vs VBP), and prepositions vs particles (run up a hill vs run up a bill).
   • Marcus et al., 1993, Building a Large Annotated Corpus of English (Computational Linguistics), §2.3