Rethinking The Corpus: Moving towards Dynamic Linguistic Resources

摘要

The corpus is an invaluable resource in Spoken and Natural Language Processing. Consistent data sets have allowed for empirical evaluation of competing algorithms. The sharing of high-quality annotated linguistic data has enabled participation and experimentation by a wide range of researchers. However, despite dubbing these annotations as "gold-standard", many corpora contain labeling errors and idiosyncrasies. The current view of the corpus as a static resource makes correction of errors and other modifications prohibitively difficult. In this paper, a perspective of the corpus as dynamically changing is advanced. We highlight the problems of the static view of the corpus through case studies of the Penn Treebank, Switchboard, Hub-4 and Boston University Radio News Corpus. We propose the use of version control software as a mechanism to facilitate this dynamic view.