Automated Detection of Local Coherence in Short Argumentative Essays Based on Centering Theory

We describe in this paper an automated method for assessing local coherence in short argumentative essays. We use ideas from Centering Theory to measure local coherence of essays’ paragraphs and compare it to human judgments on one analytical feature of essay quality called Continuity. Paragraphs which correspond to a discourse segment in our work and which are dominated by one prominent concept were deemed locally coherent according to Centering Theory. A dominance measure was proposed based on which local coherence was judged. Results on a corpus of 184 argumentative essays showed promising results. Our findings also suggest that focusing on nominal subject for detecting candidate concepts for a discourse segment’s central concept is sufficient, which confirms previous findings. Compared to previous approaches to assessing local discourse coherence in essays, our method is fully automated.