Mapping Multichannel Documents to Attentive Tasks: Ensuring Information Gain and Detecting Failure

Our approach of process-driven document analysis (DA) aims at supporting enterprises in managing the complexity of multiple input channels. Within this approach, we proposed earlier to map each incoming document to its corresponding task context - denominated as Attentive Task (AT) - by applying search with Dempster-Shafer theory. In this paper, we extend the search algorithm with methods from machine learning for addressing the challenges in real enterprise domains: (1) information gain trees for optimizing initial evidence selection and (2) five strategies for detecting search failures. We evaluate all proposed methods on a corpus from a financial institution and give an overview on how the approach enables automation services in multichannel management.