Enhancing Attentive Task Search with Information Gain Trees and Failure Detection Strategies

Kristin Stamm; Andreas Dengel
In: Joaquim Filipe; Ana Fred (Hrsg.). Proceedings of the 5th International Conference on Agents and Artificial Intelligence. International Conference on Agents and Artificial Intelligence (ICAART-13), February 15-18, Barcelona, Spain, Pages 81-90, Vol. 2, ISBN 978-989-8565-39-6, SCITEPRESS – Science and Technology Publications, 2/2013.


Enterprises today are challenged by managing requests arriving through all communication channels. To support service employees in better and faster understanding incoming documents, we developed the approach of process-driven document analysis (DA). We introduced the structure Attentive Task (AT) to formalize information expectations toward an incoming document. To map the documents to the corresponding AT, we previously developed a novel search approach that uses DA results as evidences for prioritizing all AT. With this approach, we consider numerous task instances including their context instead of a few process classes. The application of AT search in enterprises raises two challenges: (1) Complex domains require a structured selection of well performing evidence types, (2) a failure detection method is needed for handling a substantial part of incoming documents that cannot be related to any AT. Here, we apply methods from machine learning to meet these requirements. We learn and apply information gain trees for structuring and optimizing evidence selection. We propose five strategies for detecting documents without ATs. We evaluate the suggested methods with two processes of a financial institution.

Weitere Links