Process Model Generation from Natural Language Text - Advanced Information Systems Engineering - page 492

of sentence structures, which rely on keywords like “VALIDATES THAT” or
“MEANWHILE”. Therefore, it can hardly be used in the initial process defini-
tion phase as it would require rewriting of process-relevant documents.
The University of Rio de Janeiro focuses on the derivation of BPMN models
from group stories provided in Portuguese [23]. The approach was tested with
a course enrollment process modeled by students. The examples in their paper
show that process models can be created successfully, but a couple of their ex-
hibits show that syntactical problems can occur, e.g. implicit conditions, which
we explicitly tackle with our approach. The R-BPD toolkit from the Univer-
sity of Wollongong uses a syntax parser to identify verb-object phrases [21]. It
also identifies textual patterns like “If < condition/event > ,[then] < action > ”[20].
The result are rather BPMN snippets than fully connected models. Nevertheless,
this toolkit is able to take existing models into account for cross validation.
A fifth approach is the one of Policy-Driven Process Mapping [37]. First, a
procedure was developed which creates a BPMN diagram, given that data items,
tasks, resources (actors), and constraints are identified in an input text document.
Although the approach does not require a process description to be sequential,
it does not support Pools, Data Objects, and Gateways other than an exclusive
split. Furthermore, user-interaction is required at several stages.
The approach by Sinha et al. builds on a linguistic analysis pipeline [22,38].
First, text is preprocessed with a part-of-speech tagger. Next, words are anno-
tated with dictionary concepts, which classify verbs using a domain ontology.
Then, an anaphora resolution algorithm and a context annotator are applied.
The resulting information is then transferred to a Use Case Description meta-
model and later into a BPMN process model. The dictionary concepts, which
are a vital part of their approach, rely on a domain ontology which has to be
hand-crafted. This imposes a manual effort when transferring the system to other
types of texts or languages. Instead, our approach builds on the free WordNet
and FrameNet lexical databases, which are available for different languages.
6Conluon
In this paper, we presented an automatic approach to generate BPMN mod-
els from natural language text. We have combined existing tools from natural
language processing in an innovative way and augmented them with a suitable
anaphora resolution mechanism. The evaluation of our technique shows that for
a set of 47 text-model pairs from industry and textbooks, we are able to generate
on average 77% of the models correctly.
Despite these encouraging results, we still require empirical user studies. Such
studies should investigate whether humans find the generated models useful and
easy to adapt towards a fully accurate model. Furthermore, our system is able to
read process descriptions consisting of full sentences. Furthermore, we assumed
the description to be sequential and to contain no questions and little process-
irrelevant information. Another prerequisite is that the text is grammatically
correct and constituent. Thus, the parsing of structured input, like tables or