A Further Step Towards Automatic DomainModelling by Relevant Information ExtractionLars Krupp; Gernot Bahle; Agnes Gruenerbl; Paul Lukowicz
In: 2020 EEE International Conference on Pervasive Computing and Communications. IEEE International Conference on Pervasive Computing and Communications (PerCom), In 16th IEEE Workshop on Context Modeling and Reasoning (CoMoRea), located at 17th IEEE International Conference on Pervasive Computing and Communications, PerCom'20, Austin, Texas, USA, IEEE, 2020.
A still unsolved issue in human activity recognition,as it is being used in many smart systems, is the availabilityof labeled training data or, in other words, information aboutwhat sensors actually are recording. A possible solution to thisproblem is to build detailed semantic domain models specifying,in different detail, complex compound activities. Such modelswould allow retrieving the required information without thenecessity of time-consuming labeling. On our way to develop amethod to leverage text-based domain descriptions automaticallyto build such domain models, we previously introduced a methodto extract domain-relevant information from texts. However,this method still included the requirement to hand-craft so-called “regular expressions,” which have to be adjusted or re-built for different languages and different styles. Even thoughonly done once for a language or style, the necessity of hand-crafting regular expressions (and their fine-tuning by experts)still requires extensive work. In this paper, we present the nextstep to automatize information extraction by substituting theregular expressions with an automated neural network. All stepsin this method are now fully automatic and do not require anyhand-crafting. The performance of this new method is equal tothe performance of the regular expressions method before (70%precision and recall).