Incorporating Coreference to Automatic Evaluation of Coherence in Essays

摘要

The paper contributes to the task of automated evaluation of surface coherence. It introduces a coreference-related extension to the EVALD applications, which aim at evaluating essays produced by native and non-native students learning Czech. Having successfully employed the coreference resolver and coreference-related features, our system outperforms the original EVALD approaches by up to 8% points. The paper also introduces a dataset for non-native speakers' evaluation, which was collected from multiple corpora and the parts with missing annotation of coherence grade were manually judged. The resulting corpora contains sufficient number of examples for each of the grading levels.