Robust Argumentative Zoning for Sensemaking in Scholarly Documents

We present an automated approach to classify sentences of scholarly work with respect to their rhetorical function. While previous work that achieves this task of argumentative zoning requires richly annotated input, our approach is robust to noise and can process raw text. Even in cases where the input has noise (as is obtained from optical character recognition or text extraction from PDF files), our robust classifier is largely accurate. We perform an in-depth study of our system both with clean and noisy inputs. We also give preliminary results from in situ acceptability testing when the classifier is embedded within a digital library reading environment.