Extracting Declarative Process Models from Natural Language

Abstract

Process models are an important means to capture information on organizational operations and often represent the starting point for process analysis and improvement. Since the manual elicitation and creation of process models is a timeintensive endeavor, a variety of techniques have been developed that automatically derive process models from textual process descriptions. However, these techniques, so far, only focus on the extraction of traditional, imperative process models. The extraction of declarative process models, which allow to effectively capture complex process behavior in a compact fashion, has not been addressed. In this paper we close this gap by presenting the first automated approach for the extraction of declarative process models from natural language. To achieve this, we developed tailored Natural Language Processing techniques that identify activities and their inter-relations from textual constraint descriptions. A quantitative evaluation shows that our approach is able to generate constraints that closely resemble those established by humans. Therefore, our approach provides automated support for an otherwise tedious and complex manual endeavor.

Extracting Annotations from Textual Descriptions of Processes

“…As we will see in the evaluation section, our approach is significantly more accurate with respect to the state-of-the-art in the extraction of the main process elements. Likewise, we have incorporated as well the patterns from [16], and a similar outcome is reported in the experiments. Moreover, we believe that our ideas can be easily applied in the scope of legal documents, proposed in [3,19].…”

“…For the sake of space, we only report here the related work that focuses on the extraction of process knowledge from textual descriptions [1,7,16] or legal documents [3,19], or the work that considers textual annotations in the scope of BPM [10,12].…”

“…Additionally, we incorporated the approach used in [16] to extract relations Response, Precedence, and Succession based on the mandatory aspect of each activity: Given a sentence stating that activity A happens before activity B, if both A and B are mandatory, the relation is marked as Succession. If B is mandatory but A is not, the result is a Response relation.…”

“…This section presents experiments evaluating the performance of the proposed approach 9 to automatically extract ATDP from text. We will report two different experiments for two types of evaluation: we will compare activity extraction with a baseline based on [6], and relation extraction with the recent approach [16].…”

“…The ubiquity of textual descriptions of processes has caused recent research in proposing mechanisms to make actionable this information source, e.g., the discovery of formal process models [7,16], the alignment between structured and non-structured process information [13,17], or the use of annotations and natural language techniques to encompass process information [12,10].…”