Data-Driven Annotation of Textual Process Descriptions Based on Formal Meaning Representations

Titelangaben

Ackermann, Lars ; Neuberger, Julian ; Jablonski, Stefan:
Data-Driven Annotation of Textual Process Descriptions Based on Formal Meaning Representations.
In: la Rosa, Marcello ; Sadiq, Shazia ; Teniente, Ernest (Hrsg.): Advanced Information Systems Engineering : CAiSE 2021, Proceedings. - Cham : Springer , 2021 . - S. 75-90 . - (Lecture Notes in Computer Science ; 12751 ) (Information Systems and Applications, incl. Internet/Web, and HCI ; 12751)
ISBN 978-3-030-79382-1
DOI: https://doi.org/10.1007/978-3-030-79382-1_5

Volltext

Link zum Volltext (externe URL):

Abstract

Business process management encompasses a variety of tasks that can be solved system-aided but usually require formal process representations, i.e. process models. However, it requires a significant effort to learn a formal process modeling language like, for instance, BPMN. Among others, this is one reason why companies often still stick to informal textual process descriptions. However, in contrast to formal models, information from natural language text usually cannot be automatically processed by algorithms. Hence, recent research also focuses on annotated textual process descriptions to make text machine processable.

While still human-readable, they additionally contain annotations following a formal scheme. Thus, they also enable automated processing by, for instance, formal reasoning and simulation. State-of-the-art techniques for automatically annotating textual process descriptions are either based on hand-crafted rule sets or artificial neural networks. Maintaining complex rule sets requires a significant manual effort and the approaches using neural networks suffer from rather low result quality. In this paper we present an approach based on Semantic Parsing and Graph Convolutional Networks that avoids manually defined rules and provides significantly better results than existing techniques based on neural networks. A comprehensive evaluation using multiple data sets from both academia and industry shows encouraging results and differentiates between several applied text features.