Information Technology Reference
In-Depth Information
type “Exception”. If the World Model contains a flow of this type, an exception
intermediate event is attached to the task which serves as a source and this In-
termediate Event is connected to the target instead of the node itself. We then
skip dummy actions, which were inserted between gateways directly following
each other.
Step four is concerned with open ends. So far, no Start and End Events
were created. This is accomplished in this step. The procedure is also straight
forward. We create a preceding Start event to all Tasks which do not have any
predecessors (in-flow = 0) and succeeding End Events to all Tasks which do
not have any successors (out-flow = 0). Additionally, Gateways whose in- and
out-flow is one receive an additional branch ending in an End Event.
The last step in the model creation phase handles Meta-Activities (Issue 3.3).
We search and remove redundant nodes directly adjacent to Start or End Events.
This is required as several texts contain sentences like “[...] the process flow at
the customer also ends.” or “The process of “winning” a new customer ends
here.” If such sentences are not filtered, we might find tasks labeled “process
ends” right in front of an end event or “start workflow” following a start event.
We remove nodes whose verb is contained in the hypernym tree of “end” or
“start” in WordNet if they are adjacent to a Start or End Event.
The execution of these five steps yields a full BPMN model. As the elements of
this model do not contain any position information yet, our generation procedure
concludes with an automated layout algorithm. We utilize a simple grid layout
approach similar to [26], enhanced with standard layout graph layout algorithms
as Sugiyama [27] and the topology-shape-metric approach [28]. For the example
text of the claims handling process from Section 2 we generated the model given
in Figure 5. The question of how far this result can be considered to be accurate
is discussed in the following section.
4 Evaluation of Generated Process Models
For the validation of our approach, we collected a test data set consisting of
47 of those text-model pairs, each including a textual process description and
a corresponding BPMN models created by a human modeler. Different sources
from research and practice were incorporated into our test data set: Academic
(15 models), Industry (9 models), Textbook (9 models), and Public Sector (14
models), see Table 2. While the academic pairs were provided by our university
partners, the industry models are taken from two main sources. First, we gath-
ered four models from the websites of three BPM tool vendors, namely Active
VOS, Oracle, and BizAgi. Four models stem from training material of inubit
AG, another one from a German BPM practitioner, and further ones from two
BPMN textbooks [29,9]. Finally, we included the definition of switch processes
of the
Federal Network Agency
of Germany in its semi-structured tabular format
and the corresponding model.
To avoid unintended effects while parsing, minor corrections were applied
to the texts. Some models were translated, some were converted from other
