Part-Of-Speech in Historical Corpora: Tagger Evaluation and Ensemble Systems on ARCHER

Abstract

Tagger accuracy deteriorates when applied to texts different from the training corpus, e.g. with respect to register or time period. On historical data, accuracy can drop to and below 90%. We are tagging and parsing ARCHER, a historical corpus sampled from British and American texts from 1600-1999. We improve tagging accuracy by (1) using a version of the corpus that has been automatically mapped to PDE spelling with VARD, (2) by combining several part-of-speech taggers in an ensemble system – which improves tagging by about 1% over CLAWS and 2% over Tree-Tagger, and (3) by using a small amount of human intervention – which allows us to reach 98% accuracy from 1700 on.