Error Detection for Treebank Validation

Abstract

This paper describes an error detection mechanism which helps in validation of dependency treebank annotation. Consistency in treebank annotation is a must for making data as errorfree as possible and for assuring the usefulness of treebank. This work is aimed at ensuring this consistency and to make the task of validation cost effective by detecting major errors induced during completely manual annotation. We evaluated our system on the Hindi dependency treebank which is currently under development. We could detect 76.63 % of errors at dependency level. Results show that our system performs well even when the training data is low. 1