Jens Knoop, Uwe Zdun (Hrsg.): Software Engineering 2016,
Lecture Notes in Informatics (LNI), Gesellschaft f
¨
ur Informatik, Bonn 2016 71
Supporting Process Model Validation through Natural
Language Generation
Henrik Leopold
1
Jan Mendling
2
Artem Polyvyanyy
3
The design and development of process-aware information systems is often supported by
specifying requirements as business process models. Although this approach is generally
accepted as an effective strategy,itremains afundamental challenge to adequately vali-
date these models giventhe diverging skill set of domain experts and system analysts. As
domain experts often do not feel confident in judging the correctness and completeness
of process models that system analysts create, the validation often has to regress to adis-
course using natural language. In order to support such adiscourse appropriately,so-called
verbalization techniques have been defined for different types of conceptual models. How-
ever,there is currently no sophisticated technique available that is capable of generating
natural-looking text from process models. The reason whyaproper process model verbal-
ization technique is still missing might be aresult of the difficulty to meet this challenge.
Aprocess model verbalization technique has to serialize the non-sequential structure of
aprocess model into sequential, yet execution-order preserving, text. In addition, it must
be capable of analyzing the short and grammatically varying labels of process model el-
ements and of annotating them with their semantic components likeaction or business
object. Furthermore, the verbalization technique needs to handle optionality of certain
pieces of information. In the paper [LMP14], we address this research gapand propose a
technique for generating natural language texts from business process models.
WordNet
Stanford
Tagger
BPMN
Process
Model
Linguistic
Information
Extraction
Annotated
RPST
Generation
Text
Structuring
DSynT-
Message
Generation
Message
Refinement
Text Planning
Sentence Planning
Realization
RealPro-
Realizer
Natural
Language
Text
Abb.1:Architecture of our Natural Language Generation System
The architecture of our text generation technique is building on the traditional pipeline
concept from natural language generation systems. The basic rationale of the technique
is to utilize the existing information from the model to generate atext. In order to derive
1
VU University Amsterdam, De Boelelaan 1081, 1081HV Amsterdam, The Netherlands, h.leopold@vu.nl
2
WU Vienna, Welthandelsplatz 1, 1020 Vienna, Austria, jan.mendling@wu.ac.at
3
Queensland University of Technology,Brisbane, QLD 4001, Australia
72 Henrik Leopold et al.
asequence of sentences, we linearize the model via the creation of atree structure. In
particular,the text generation technique comprises six components (see Figure 1):
1. Linguistic Information Extraction:Extraction of linguistic components from the
process model element labels.
2. Annotated RPST Generation:Linearization of process model through the generation
of atree structure. In addition, each node is annotated with the linguistic information
from the previous step.
3. Text Structuring:Application of text structuring techniques, such as the insertion of
paragraphs and bullet points, based on the computed tree structure.
4. DSynT-Message Generation:Generation of an intermediate linguistic message struc-
ture for each node of the tree. This component represents the core of the generation
technique.
5. Message Refinement:Refinement of the generated messages through aggregation or
the introduction of referring expressions and discourse markers.
6. RealPro-Realizer:Transformation of intermediate message structures to grammati-
cally correct sentences.
To demonstrate the capability of the proposed technique for generating natural language
texts from process models, we conducted atwo-step evaluation. First, we applied our tech-
nique to real-world process models and investigated howthe generated texts compare to
textual descriptions created by humans. Second, we studied in howfar humans are capable
of making sense of the generated texts. To this end, we asked humans to transform the gen-
erated texts back into process models. The first evaluation step showed that the generated
texts convey the model semantics in amore compact and also syntactically less complex
manner.Due to the design of the technique, the generated texts are closer to the model and
describe the model content and control flowexplicitly.The second evaluation step demon-
strated that the generated texts are very informative and can successfully be interpreted by
humans.
Based on our findings, we conclude that the proposed text generation technique has the po-
tential to facilitate the validation discourse between domain experts and process analysts.
First, the generated texts support domain experts in understanding the details of process
models even if theyare not familiar with process modeling. Second, the text generation
may also train domain experts in reading and interpreting process models. As long as text
and model are presented together,readers can see and learn about the connection between
model and text. Thus, their overall familiarity with process models can be expected to
increase in the long term.
References
[LMP14] Leopold, Henrik; Mendling, Jan; Polyvyanyy,Artem: Supporting process model valida-
tion through natural language generation. Software Engineering, IEEE Transactions on,
40(8):818–840, 2014.