Evaluating Coherence of Essays using Sentence-similarity Networks

The main weakness of automated essay evaluation systems is their predominant focus on vocabulary and text syntax, while consideration of text semantics is often neglected. In this work, we propose several new attributes for measuring coherence and consistency of essays that are based on a network representation of essays. In this representation, nodes represent sentences and links reflect similarity between them. We evaluated the proposed attributes on a benchmark dataset showing that their integration into a state-of-the-art system for essay evaluation indicates a potential for improvement of predictive performance.