An Automated and Intelligent Interface Embracing Process Awareness
into User Workspace
Minh Khoi Nguyen
1 a
, Hanh Nhi Tran
1,2 b
, Ileana Ober
1 c
and Razan Abualsaud
1 d
1
Institut de Recherche en Informatique de Toulouse (IRIT), Toulouse, France
2
Military Academy of Saint-Cyr Co
¨
etquidan (AMSCC), CReC Saint-Cyr, France
Keywords:
Artificial Intelligence, Process Management System, Process Monitoring, Collaborative Process.
Abstract:
This paper presents an AI-augmented framework for automated and intelligent process monitoring, addressing
the inefficiencies of manual progress reporting in Process Management Systems (PMS), which leads to poten-
tial inaccuracies and consumes valuable user time. Our research proposes a novel solution that bridges users’
workspaces and PMS, enabling automatic progress reporting based on users’ actions within their preferred
tools. The core innovation of our framework pMage lies in employing Artificial Intelligence (AI) techniques
to analyze and interpret sequences of user actions, translating them into accurate task progress updates, which
significantly reduce manual input and enhance the accuracy of the reporting, thus making the integration of a
PMS smoother and more effective. We demonstrate our framework’s applicability through a case study that
uses pMage to monitor a brake system manufacturing process with our prototype. As a smart interface, pMage
provides a no-code solution to connect a wide range of user applications to various PMS via their respective
APIs. This versatility ensures broad applicability across different organizational contexts and toolsets. Our
AI-augmented framework offers a more reliable, efficient, and user-friendly approach than existing monitoring
methods.
1 INTRODUCTION
In contemporary organizational environments, Pro-
cess Management Systems (PMS) play a crucial role
in monitoring and managing the progress of various
tasks and projects. Traditionally, PMSs rely heavily
on user inputs to track work progress, which presents
significant challenges. Typically, users perform tasks
using their preferred tools within their workspace and
subsequently report their progress to the PMS manu-
ally. This method is fraught with issues, including the
risk of inaccurate reporting and the considerable time
investment required from users to log their activities.
Our research addresses this critical problem by
proposing an innovative intermediate framework that
acts as an intelligent interface between PMS and pro-
cess participants’ diverse workspaces. The frame-
work’s primary objective is to enable automatic and
real-time process progress reporting by capturing and
a
https://orcid.org/0000-0001-5412-7259
b
https://orcid.org/0000-0003-0868-1253
c
https://orcid.org/0000-0001-9338-8187
d
https://orcid.org/0009-0007-4895-1242
interpreting users’ actions within their preferred tools.
This seamless integration eliminates the need for
manual updates, enhances tracking accuracy, and pro-
vides real-time visibility into process execution. By
bridging the gap between user activities and process
management, our framework aims to significantly im-
prove the efficiency and timeliness of process mon-
itoring, ultimately leading to more effective process
management and decision-making.
The rest of this paper is organized as follows: Sec-
tion 2 introduces the research questions and the con-
tribution of this paper. The related works and current
approaches are introduced in Section 3. Section 4 pro-
vides an overview of the core of our solution pMage.
Section 5 focuses on our AI-driven solutions for of-
fering a smart interface enhancement pMage that sup-
ports automated and intelligent process monitoring.
Section 6 evaluates the proposed framework in the
context of the case study on the process of manufac-
turing a brake system for a self-driving car, while Sec-
tion 7 discusses the strengths and weaknesses of our
approach and suggests directions for future research.
Nguyen, M. K., Tran, H. N., Ober, I. and Abualsaud, R.
An Automated and Intelligent Interface Embracing Process Awareness into User Workspace.
DOI: 10.5220/0013370200003896
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 13th International Conference on Model-Based Software and Systems Engineering (MODELSWARD 2025), pages 97-108
ISBN: 978-989-758-729-0; ISSN: 2184-4348
Proceedings Copyright © 2025 by SCITEPRESS – Science and Technology Publications, Lda.
97

2 RESEARCH QUESTIONS
Defining a framework that allows to seamlessly inte-
grate user workspace and process execution raises the
following research questions:
Figure 1: The intermediate framework to integrate end-user
workspaces evoking diverse actions with PMSs.
RQ1. How to manage and exploit the heterogene-
ity of tools and PMS ? As shown in Fig. 1, the
framework must be versatile enough to inter-
face with various software environments (e.g.
AutoCAD, GitHub, SIMATIC, etc.), each with
its own data structures, APIs, and communica-
tion protocols. Balancing the need for a stan-
dardized approach with the flexibility to ac-
commodate this diversity of tools is a complex
undertaking.
RQ2. How to accurately extract high-level task
completion from the numerous low-level ac-
tions users perform within their workspace.
For example in Fig. 1, in a particular process
execution, which task corresponds to the ac-
tion of start job in SIMATIC? This requires
sophisticated interpretation of user behaviors
across various applications having their unique
interface and functionality.
In our previous work (Nguyen et al., 2024), we
developed a metamodel that abstracts tool families,
enabling a homogeneous connection mechanism via
tool APIs to manage tool heterogeneity. Inspired
by the vision of (Dumas et al., 2023) to integrate
more AI techniques into a business PMS to activate
the intelligent adaptability and self-improvement of a
business process, we also implemented an ontology-
based mechanism to map low-level application ac-
tions to high-level tasks to solve the second research
question. This ontology was built semi-automatically
by extracting information from past projects (pro-
cess names, task names, artifacts, actor roles, etc.)
using an AI-augmented process mining framework
(Nguyen et al., 2023). However, the extracted in-
formation from historical project data is often limited
and potentially biased, providing inadequate contex-
tual information on user actions, and leading to inac-
curacies in mapping these actions to high-level tasks.
This paper presents two major contributions to au-
tomated process monitoring, providing machine in-
telligence to improve human productivity (Van der
Aalst, 2021).
C1. We introduce an innovative procedure for dis-
covering appropriate connection configura-
tions among heterogeneous tools by leverag-
ing historical connection data. This contribu-
tion enhances the framework’s ability to adapt to
diverse toolsets and organizational contexts, fur-
ther streamlining the integration process and im-
proving the overall flexibility of our framework.
C2. We replace the previous ontology-based solu-
tion with a more sophisticated and powerful
AI model to infer task progression from user ac-
tions with higher precision. This AI model is
trained using enriched data from past projects
and web sources, significantly expanding the
knowledge base and reducing potential biases.
These advancements represent a substantial leap
forward in the accuracy, adaptability, and efficiency
of automated process monitoring.
3 RELATED WORKS
Various approaches aim to provide the user with
a smooth embrace of process management. While
(Due
˜
nas et al., 2018) proposed gathering informa-
tion on the work progress of the project, (Delgado
et al., 2016) introduced a central generic portal to en-
gage different PMSs. Furthermore, some approaches
propose a platform that allows monitoring the user
work progress and updating the state of the process
instance simultaneously (Baresi et al., 2017; Baresi
et al., 2016).
To ensure users can continuously and transpar-
ently utilize a familiar PMS, (Delgado et al., 2016)
proposed a generic PMS user portal designed to in-
tegrate with various concrete PMSs. This portal is
built on a unified data model and a generic process
engine API, enabling it to offer functions tailored to
specific business contexts while minimizing disrup-
tions caused by changes in the underlying PMS. The
portal consists of two layers: a presentation layer,
which defines the generic user interface, and an ac-
cess layer, which facilitates the connection between
the portal and specific process engines (e.g., Bonita,
Activiti, and Bizagi). Although the proposed portal
provides comprehensive PMS functionality and re-
MODELSWARD 2025 - 13th International Conference on Model-Based Software and Systems Engineering
98

duces the time required for training and usage, it does
not fully free users from task reporting, limiting their
ability to focus entirely on their primary working ap-
plications.
The need for an efficient mechanism to align spe-
cific work with the process model in a PMS has
spurred various research efforts. (Cohn and Hull,
2009) highlights that the state of a process instance
is inherently tied to the state of the business entities
(or artifacts) involved and produced during its execu-
tion. Each artifact is characterized by an information
schema and a life cycle, which defines how it evolves
throughout the process (Dumas, 2011). Building
on this concept, (Baresi et al., 2017) introduces
an artifact-driven process monitoring platform called
mArtifact, capable of flagging affected activities dur-
ing violations without interrupting the monitoring
process. This platform proves particularly effective
for collaborative, multi-site environments. Similarly,
(Baresi et al., 2016) proposes leveraging Smart Ob-
jects to monitor object states in cross-organization
business processes. This approach, grounded in the
GSM (Guard-Stage-Milestone) framework, enables
the comparison of process instances against their
models to detect control-flow violations and activity
faults. To address storage limitations, Smart Objects
only retain information about the activities that users
intend to monitor.
To facilitate the management of the software de-
velopment process, (Due
˜
nas et al., 2018) introduced
Perceval, a platform designed to automatically and
incrementally collect data from various tools related
to open-source development and present it through
a centralized dashboard. Perceval is intended to
be highly extensible, enabling cross-cutting analy-
sis and providing incremental updates—particularly
valuable for analyzing large software projects. How-
ever, Perceval does not handle data storage or analy-
sis itself, so these tasks are delegated to other tools.
Moreover, it lacks a concrete method for integration
with specific PMSs, offering no direct support for
end-users in task reporting.
These approaches rely on built-in platforms to
manage processes, which require significant addi-
tional effort to customize for specific needs, domains,
and contexts. While they offer the advantage of being
adaptable to a user’s unique requirements, they incur
substantial costs in terms of the time and effort needed
for development, testing, and deployment.
On the other hand, limited research has focused on
leveraging machine learning or AI to enhance process
awareness within user workspaces. (Weinzierl et al.,
2024) highlights numerous studies that integrate and
innovate machine learning and AI techniques across
various aspects of business process management, in-
cluding process identification, process discovery, pro-
cess analysis, process redesign, process implementa-
tion, and process monitoring. Among these, process
monitoring - most closely related to process aware-
ness - commonly assumes the availability of perfectly
structured event log data. Research in process mon-
itoring has primarily explored methods for extract-
ing features from online event logs (Leontjeva et al.,
2015), dynamically reconstructing process models,
and predicting process execution outcomes (Ever-
mann et al., 2017; Metzger et al., 2019). In contrast,
our proposed framework tackles the critical challenge
of directly capturing event logs that reflect user be-
havior during the process, specifically within the user
workspace.
4 pMage OVERVIEW
Figure 2: Architecture of pMage.
The architecture of pMage (illustrated in Fig. 2) has
three main components: (1) an Intelligent Connector
– that archives integration configurations and estab-
lishes connections between end-user applications and
PMSs, (2) an Intelligent Monitor – that reports the
progress of the impacted high-level tasks in the PMS
from low-level user actions in their applications, and
(3) a Data Warehouse – that keeps track of task exe-
cutions for further analysis and learning.
The core component of our framework is the Intel-
ligent Monitor, which acts as a bridge between user
applications and the PMS to interpret low-level user
actions to high-level task progress within the PMS. It
is responsible for the following process ensuring that
each user action is accurately reflected in the PMS in
real-time without requiring manual updates:
• Event Triggering: as shown in Fig. 3, when a
user performs an action in their application, the
application triggers an event e1 within the user’s
workspace.
• Action Detection: The Intelligent Monitor detects
this triggered event e1, signaling that a user action
An Automated and Intelligent Interface Embracing Process Awareness into User Workspace
99

has occurred.
• Action Interpretation: To interpret this action, the
Intelligent Monitor consults the Action Linkage
table (c.f. Fig. 3). This table is a crucial compo-
nent of our framework, storing precise mappings
between low-level user actions in the workspace
and corresponding high-level task actions in the
PMS (mapping from App Event e1 to PMS Event
e2 on task t).
• Task Identification: By referencing the Action
Linkage table, the Intelligent Monitor identifies
the specific task t associated with the user’s ac-
tion.
• Progress Determination: Based on the mapping,
the Intelligent Monitor determines the progress
made on the identified task t and then triggers the
corresponding PMS event e2 to update the task’s
progress accordingly.
Figure 3: Action Linkage table provides the mapping ref-
erences enabling identifying the impacted high-level task
from the low-level user action.
The necessity for an accurate and comprehensive
Action Linkage table comes from its role of enabling
our framework to automatically interpret user actions
and translate them into meaningful task progress up-
dates. Yet, the creation of the Action Linkage table
presents significant challenges and opportunities for
innovation.
The Action Linkage table handles four aspects
(columns in Fig. 3), each playing a specific role in
the action-to-task mapping process. First, the event
e1, triggered by a low-level action performed by the
user within their workspace. It serves as the starting
point for the mapping process, capturing the initial
user interaction that needs to be interpreted. Second,
the contextual information of the user action is essen-
tial for identifying the related task t in the PMS af-
fected by the user’s action. Furthermore, the event e2
triggers an action in the PMS to update the relevant
task t.
Traditionally, filling in such a table would require
users to manually define the mappings based on their
understanding of the tools and tasks involved in the
process. Although this method can be precise, it is
time-consuming and prone to human error, particu-
larly when managing numerous actions across multi-
ple tools. Additionally, as processes evolve and new
tools are introduced, manually updating the table be-
comes increasingly difficult. This approach also de-
mands extensive expertise in tools and processes con-
currently, which may not always be readily accessi-
ble.
To address these challenges and facilitate the use
of our framework, our objective is to generate the Ac-
tion Linkage table to automate the mapping process,
thereby reducing the time and effort required from
users. While the data for the application event e1 and
PMS event e2 can be extracted from the configura-
tions of the user tool and the PMS respectively, filling
the second and the fourth columns is more compli-
cated. The second column containing contextual in-
formation about the user action is crucial for accurate
task identification for filling the fourth column. This
context is vital because the same user action can be
associated with different tasks depending on the cir-
cumstances in which it is performed. For instance,
saving a file might update a document creation task
in one context, but signal the completion of a review
process in another.
Determining the contextual information and cor-
responding process tasks requires an expert with deep
knowledge of both the tool and the process. To over-
come this dependency and enhance efficiency, we
have developed an AI model that learns to provide
this contextual information automatically (see Section
5). This AI-driven approach aims to replicate the nu-
anced understanding of an expert, enabling the frame-
work to accurately discern the appropriate context for
each user action and associate it with the correct task.
By structuring the Action Linkage table in this man-
ner and leveraging AI for context determination, our
framework efficiently and accurately maps user ac-
tions to task updates. This approach significantly re-
duces the need for manual configuration while main-
taining the necessary context awareness for precise
task mapping, even when identical user actions cor-
respond to different tasks based on their context.
5 DEVELOPMENT OF
INTELLIGENT CONNECTOR
AND INTELLIGENT MONITOR
This section details the two main advancements of
our framework, which address the research questions
MODELSWARD 2025 - 13th International Conference on Model-Based Software and Systems Engineering
100

RQ1 and RQ2. First, we present the Intelligent Con-
nector, which enables more efficient connection es-
tablishment (Sec. 5.1). Subsequently, we describe the
Intelligent Monitor, which enables automated action
interpretation and intelligent task monitoring (Sec.
5.2).
5.1 Intelligent Connector
Integrating applications and PMSs to monitor process
execution within a project requires users to input es-
sential information, such as the application’s name,
project directory location, and login username. This
task becomes tedious and time-consuming when re-
peated for multiple projects.
To streamline this process, we leverage past con-
nection data from the Data Warehouse to simplify
new connection setups. This approach particularly
benefits new users by suggesting common applica-
tions and PMS configurations for specific domains.
These suggestions help users identify appropriate
tools and provide necessary configuration and event
profiles for intelligent monitoring.
Figure 4: Workflow of Intelligent Connector.
The detailed method for implementing this ap-
proach is illustrated in Fig. 4 and Algo. 1. The main
challenge is identifying the key information needed
for effective configuration suggestions. Users expect
an automatic field population based on the project
name. Thus, the core of our suggestion process re-
lies on keyword correlation between the new project
name and historical projects’ representative terms –
project, task, and artifact names.
Our approach involves two key phases:
P1. Collecting Historical Connections Configura-
tion
S1.1. By applying Natural Language Processing
(NLP) including tokenizing, removing punc-
tuation, stopwords, and lemmatizing, we col-
lect representative terms of past connections
from their process, task, and artifact names.
Algorithm 1: Connection Configuration Suggestion.
Input: username, processDesc, pastConnections
Output: bestConfig
forall connection ∈ pastConnections do
representTerms :=
nlp(connection.processName ∪
connection.taskNames ∪
connection.artifacts)
forall term ∈ representTerms do
relatedWords := findRelatedWords(term)
extendedTerms.add(relatedWords)
end
connection.representTerms := extendedTerms
end
candidates :=
/
0
forall connection ∈ pastConnections do
matchingWords :=
findMatchingWords(processDesc,
connection.representTerms)
candidates.put(connection, matchingWords)
end
bestConfig := findUserPastConnection(username,
candidates)
if bestConfig is null then
bestConfig := findBestCandidate(candidates)
end
S1.2. These terms are expanded with their close
meanings and synonyms to increase the poten-
tial for optimal configuration matches in P2.
P2. Suggesting Connection Configuration
S2.1. For each connection request, we extract terms
from the new project name using similar NLP
techniques.
S2.2. Compute Jaccard similarity between the new
project and historical projects. Configurations
created by the same user are prioritized for
recommendations. Otherwise, the historical
project with the highest Jaccard similarity be-
comes the suggested configuration for the new
project.
JaccardSimilarity(A, B) =
|A ∩ B|
|A ∪ B|
(1)
with A and B are the set of terms from the new
project and the historical connection respectively.
5.2 Intelligent Monitoring
As explained in Section 4, the primary challenge for
the Intelligent Monitor lies in generating the Action
Linkage table, which captures the complex mappings
between low-level user actions and high-level task up-
dates, requiring a nuanced understanding of both the
user’s workspace and the process management con-
text.
An Automated and Intelligent Interface Embracing Process Awareness into User Workspace
101

Figure 5: Workflow of fine-tuning AI model for Intelligent Monitoring.
To address the limitations of the previous solu-
tion (Nguyen et al., 2023), we leverage AI techniques
more extensively to enhance the generation of the Ac-
tion Linkage table. We propose using an AI model,
which plays the role of a domain expert, to automat-
ically generate the context information of the user’s
actions in the Action Linkage table. For this pur-
pose, we enrich a pre-trained AI model’s base knowl-
edge with domain-specific vocabulary. While pre-
trained AI models have learned a vast array of com-
mon words, enabling general information recognition,
the addition of domain-related vocabulary provides
a focused layer of understanding. This fine-tuned
AI model unveils more appropriate contextual infor-
mation, allowing for precise mapping and improved
comprehension within the target domain. By learning
from a broader, more diverse set of data and exam-
ples, the AI model generates more reliable and stable
context information for the Action Linkage of each
connection user application-PMS within the context
of a given process.
However, implementing an AI-based solution for
generating the Action Linkage table has its own chal-
lenges. Gathering a sufficient and diverse dataset of
user actions and corresponding task progress for train-
ing the AI model is a significant hurdle. Developing
algorithms that can understand the context of actions
across different tools and processes adds another layer
of complexity. Creating a model that can generalize
well across various domains and types of processes is
crucial for the framework’s versatility. Finally, ensur-
ing that the AI-generated mappings are accurate and
reliable for real-world use is paramount.
Figure 5 presents the three key proposed phases
for developing Intelligent Monitoring, along with
their detailed steps and the illustrated results from the
experiment described in Section 6:
P1. Build Domain Vocabulary: Learning domain
vocabulary from both domain-related web pages
and past projects has (1) greater accessibility to
up-to-date terms, (2) a more diverse and less bi-
ased vocabulary, and (3) a balanced integration
of user behaviors and domain knowledge. We
employ the following process to construct the
domain vocabulary:
S1.1. Search for domain-related web pages using
keywords.
S1.2. Extract content from (1) the collected domain-
related web pages, and (2) historical project
connections.
S1.3. Pre-process the extracted terms using the
same NLP techniques as for the workflow of
Intelligent Connector.
S1.4. Compute and rank the importance of each
term by calculating its TF-IDF value across
all web page content. TF-IDF (Term Fre-
quency Inverse Document Frequency) mea-
sures the relevance of a word within a cor-
pus. The importance increases proportionally
to the number of times a term t appears in a
web page p (N(t)
p
) relative to the total terms
in the page (N
p
), and is adjusted by the term
frequency (N(t)) across the entire corpus of N
MODELSWARD 2025 - 13th International Conference on Model-Based Software and Systems Engineering
102

web pages.
T F − IDF(t) =
∑
p
T F − IDF(t, p)
=
∑
p
N(t)
p
N
p
∗ log(
N
N(t)
)
(2)
S1.5. Select a fixed number of terms with the high-
est importance values as the domain vocabu-
lary.
P2. Prepare AI Model Training Dataset: This
phase is crucial for guiding the AI model to gen-
erate contextual information for the Action Link-
age table using domain-specific terms based on
the task names.
S2.1. For each term in the domain vocabulary, we
use GloVe (Global Vectors for Word Repre-
sentation) (Pennington et al., 2014) to identify
synonyms and related words.
S2.2. Extend the search to find related words at a
second level (related words of the initial re-
lated words). The ideal approach would be an
infinite loop of searches until no new words
are found, however, we limit our search to the
second level due to resource and time con-
straints. Duplicate terms are excluded, result-
ing in an enriched set of related words for each
domain term.
S2.3. Construct a training dataset to fine-tune the
pre-built AI model for text summarization.
The AI model aims to generate domain-
specific terms from the input process name.
P3. Train AI Model for Text Summarization: This
final phase focuses on training the AI model
to understand domain-specific terms and sum-
marize input text into domain-related keywords,
which will be used as contextual information in
the Action Linkage table.
S3.1. Load a pre-built Transformer model trained
on a large dataset of diverse tasks. This
pre-trained model enables transfer learning,
where knowledge from one task is adapted to
a different but related task by fine-tuning the
model on a smaller, task-specific dataset.
S3.2. Define hyper-parameters suitable for the pre-
trained AI model and the specialized text sum-
marization task (e.g. Table 2), which involves
balancing our computing resources with rec-
ommendations from other studies on the AI
model and the specialized task.
S3.3. Train the AI model with the defined hyper-
parameters using the dataset created in P2:
the general knowledge terms as the input and
the domain-related terms as the targeted out-
put. Each time the fine-tuned AI model meets
a text, it recognizes the general knowledge
terms and returns the domain-related terms to
generate better contextual information for the
Action Linkage table.
6 EVALUATION
The evaluation aims to validate the proposed intelli-
gent components by assessing their applicability and
efficiency in suggesting new integrating connections,
detecting context information from user applications,
and scalability to handle large datasets and complex
process configurations when applying to case study
Manufacture Brake System for a Self-driving Car.
6.1 Evaluation Method
The evaluation is carried out as follows:
Case Study: we imitate the procedure of integrating
the process execution of Manufacture Brake System
for a Self-driving Car (c.f. Fig. 6) into our proposed
framework, including the step of connection estab-
lishment and setting up the Action Linkage table for
monitoring user action.
This process represents a research and develop-
ment workflow, customized for the manufacturing of
a self-driving car model. Regarding the process tasks,
the tasks Design brake system blueprint, Implement
RSDs braking activation, Connect RSD with brake
control system, and Assemble the system into a pro-
totype vehicle are performed once for each car model
individually. In contrast, the task Assemble the system
into the final version is executed multiple times on
an assembly line that handles multiple vehicles of the
same model. In terms of roles, the process involves
two key contributors: the Mechanical Engineer and
the Embedded System Developer. A notable charac-
teristic of this case study is that the process is exe-
cuted by distinct actors, each utilizing entirely differ-
ent sets of applications and project management sys-
tems (PMS):
• The Embedded System Developer uses Github -
a developer platform that allows creating, storing,
and sharing code - for Implementing remote sens-
ing devices braking activation
• The Mechanical Engineer uses AutoCAD - a gen-
eral drafting and design application to prepare
technical drawings - for designing brake system
blueprint
An Automated and Intelligent Interface Embracing Process Awareness into User Workspace
103

Figure 6: Process model of Manufacturing Brake System for a Self-driving car.
• SIMATIC WinCC V8 - a process visualization
system enabling seamless monitoring and opera-
tion of automated processes - is used by both the
Embedded System Developer and the Mechanical
Engineer for assembly tasks.
The case study highlights the challenges of man-
aging complex communication between participating
applications and PMSs. A connection is needed for
each pair of user applications and PMS. Addition-
ally, a mapping from low-level user actions to high-
level task updates is required for each connection.
The presence of numerous connections within a sin-
gle project increases both the technical complexity of
establishing these connections and the potential am-
biguity in action interpretation.
Simulation: Synthetic event logs are generated
to simulate various process definitions to examine the
proposed Intelligent Connector of the framework.
For evaluating the proposed framework perfor-
mance, we use 3 different metrics:
• Jaccard Similarity: to assess the similarity be-
tween the new project name and the suggested
historical project configuration of the Intelligent
Connector.
• ROUGE (Recall-Oriented Understudy for Gist-
ing Evaluation): a set of metrics used for eval-
uating the summaries’ quality by measuring the
overlap between the generated summary and the
original text of the Action Linkage table. We fo-
cus on ROUGE-1, and ROUGE-L:
– ROUGE-1: compares the number of matching
individual words between the summary and the
original text.
– ROUGE-L: measures the longest common
subsequence (LCS) between the generated
summary and the original text but does not re-
quire consecutive words, making it a more flex-
ible and lenient similarity measure.
• BERT Score: uses contextual embeddings from
pre-trained Transformer-based models to compute
the similarity between the generated summary and
the original text of the Action Linkage table.
We expect our framework to achieve a high simi-
larity score in the connection suggestion and the gen-
erated contextual information. The similarity scores
reflect the effectiveness of a solution and help dis-
cover the impact of preparing the training dataset for
the AI model, and the diversity of history connections
on the framework performance.
6.2 Data Sources
We have made available datasets necessary for devel-
oping the intelligent components of pMage concen-
trated on the case study.
6.2.1 Historical Connections
For the time we run our case study, this data, theoret-
ically collected by pMage, is not enough to construct
a comprehensive dataset that spans various projects
across multiple domains. Consequently, simulating
user connections is a viable alternative to real data,
forming the basis for suggesting configurations.
As outlined in Algorithm 1, the process and its
task descriptions are key elements in identifying the
appropriate configuration. We design the simulated
data with these considerations:
Seven different process definitions were sim-
ulated: Modify Test-bench Wiring, Implement E-
commerce Web Application, Manufacture Brake Sys-
tem for a Self-driving Car, Certify Car Brake, Resig-
nation Procedure, Item Ordering Process, and Apply
MODELSWARD 2025 - 13th International Conference on Model-Based Software and Systems Engineering
104

Table 1: Key aspects of the simulated historical connec-
tions.
Number of process definitions 7
Number of users 12
Number of process instances 1338
Number of applications 6
Number of connections 2000
Number of PMSs 3
for Schengen Visa. Each process involves up to two
users for its execution. Additionally, six applications
were designed and tailored for a specific process or
usable across various processes.
6.2.2 Domain Vocabulary and Training Dataset
for AI Model
We use the DuckDuckGo search engine API
(duckduckgo search) to find domain-related web
pages using the keywords manufacture car brake.
DuckDuckGo provides consistent results for all users,
avoiding personalized results based on search history
and ensuring stable and unbiased search outcomes.
We use BeautifulSoup to scrape content from the
top 50 related web pages. We also gather project,
task, and artifact names from historical projects. Us-
ing nltk (natural language toolkit), we preprocess
all collected texts by removing punctuation and stop-
words and performing lemmatization. Finally, we
rank the preprocessed texts using TfidfVectorizer
from sklearn, creating a dataset of 500 domain-
related terms for AI model training.
We discover each domain-related term its related
words using GloVe. To enhance the AI model’s capa-
bility to recognize domain-related terms amidst com-
mon terms, we include not only the related words of
the domain-related words but also the related words
of those related words. This approach results in a rich
and diverse set of domain-specific terms.
The pre-trained AI model used in our experi-
ment is Large BART from https://huggingface.co/
facebook/bart-large (Lewis et al., 2019). BART uses a
transformer-based encoder-decoder architecture with
a bidirectional encoder (like BERT (Devlin et al.,
2019)) and a left-to-right decoder (like GPT). The
pretraining task involves randomly shuffling the or-
der of the original sentences and a novel in-filling
scheme, where spans of text are replaced with a sin-
gle mask token. The hyper-parameters used for fine-
tuning BART are detailed in Table 2.
6.3 Project Specification
As described in Fig. 4, the results of the connec-
tion configuration recommendation rely on the terms
Table 2: Hyper-parameters for fine-tuning Large BART.
Learning rate 0.001
Number of training epochs 3
Training batch size 2
Evaluation batch size 2
Loss function Cross-Entropy
Weight decay 0.01
in the user’s process description. Our approach also
prioritizes suggesting the user’s historical connection
rather than from the other users. For this reason, to
examine the performance of our approach in suggest-
ing connection configuration, we test different setups:
• Project name:
– contains terms only included in the histori-
cal connections: Manufacture Brake system for
Self-driving Car
– contains some terms included in the histori-
cal connections: Brake system for new electric
Peugeot.
– does not contain any terms included in any
historical connections: Develop Anti-cheat for
FPS games
• User with (1) or without (2) historical connections
6.4 Results and Discussions
6.4.1 Connection Configuration Suggestion
Deduced from Table 3, we can suggest a connection
configuration for projects whose names are composed
entirely of terms archived in historical connections.
Although the Jaccard similarity score is relatively low
(0.067), indicating a small fraction of shared terms
compared to the total terms, this is acceptable as we
have enriched the representative terms for each his-
torical connection (see Fig. 4). Conversely, projects
with partially matching names to historical connec-
tions have a lower Jaccard similarity score of 0.025,
particularly for a different project type like Certifying
Car Brake. Terms such as Peugeot and electric con-
tribute to the lower similarity score since they are not
present in the representative terms, meanwhile the re-
maining terms are more akin to the Certify Car Brake
project than to a Manufacture Brake System for Self-
driving Car project. Projects containing entirely dif-
ferent terms from the historical connections result in
a perfect 0 similarity score, making it impossible to
suggest a connection configuration if there are no sim-
ilarities between the project’s name and the historical
connections.
The results are consistent for existing users with
historical connections and new users. The main dis-
An Automated and Intelligent Interface Embracing Process Awareness into User Workspace
105

tinction for existing users is that pMage can sug-
gest authorization information, such as login creden-
tials, passwords, or personal tokens. pMage assists
all users, including those new to the framework, in
adopting a process-aware working environment with
greater ease and efficiency.
6.4.2 Action Linkage Generation
We selected the project name Manufacture Brake Sys-
tem for Self-driving Car with its suggested configura-
tion to establish the connection between application
SIMATIC WinCC V8 and PMS jBPM. The next cru-
cial step after establishing the connection is defining
the Action Linkage table. Table 4 illustrates the gen-
erated Action Linkage table for this connection.
Table 3: Connection configuration suggestions on different
setups.
Project
name User
Configuration
suggested
Jac-
card
score
Manufac-
ture Brake
system for
Self-
driving Car
(1)
Manufacture self-driving
Car Brake,
app (SIMATIC WinCC
V8, C:\Program
Files(x86)\...),
pms (jBPM,
localhost/kie-server/...)
0.067
Brake
system for
new
electric
Peugeot
(1)
Certify Car Brake,
app (SIMATIC WinCC
V8, C:\Program
Files(x86)\...),
pms (jBPM,
localhost/kie-server/...)
0.025
Develop
Anti-cheat
for FPS
games
(1) No configuration
0.000
Manufac-
ture Brake
system for
Self-
driving Car
(2)
Manufacture self-driving
Car Brake,
app (SIMATIC WinCC
V8, C:\Program
Files(x86)\...),
pms (jBPM,
localhost/kie-server/...)
0.067
Brake
system for
new
electric
Peugeot
(2)
Certify Car Brake,
app (SIMATIC WinCC
V8, C:\Program
Files(x86)\...),
pms (jBPM,
localhost/kie-server/...)
0.025
Develop
Anti-cheat
for FPS
games
(2) No configuration
0.000
We defined the event profiles for each application
specifying
1. the method of collecting the event,
2. the location of collecting the event, and
3. the patterns within the event where contextual in-
formation should appear
Below is an example of event profiles for SIMATIC
WinCC V8, which includes two events: INFO tag
start and INFO tag complete.
” ev e n t ” : [ { ” name ” : ” INFO t a g s t a r t ” , ” me tho d ” : ”
LOG” ,
” a p i I n f o ” : ”C: \ \ P rogram F i l e s ( x86 ) \\
Si eme ns \\WinCC\\ D i agn o s e ” ,
” i m p o r t a n t ” : ” [ { t i m e }] INFO [ { t a s k } ] { t a s k
} s t a r t e d f o r ’ { a r t i f a c t } ’ by u s e r ’{
user Name } ’ . ” } ,
{” name ” : ”INFO t a g c o m p l e t e ” , ” me tho d ” : ”
LOG” ,
” a p i I n f o ” : ”C: \ \ P rogram F i l e s ( x86 ) \\
Si eme ns \\WinCC\\ D i agn o s e ” ,
” i m p o r t a n t ” : ” [ { t i m e }] INFO [ { t a s k } ] { t a s k
} c o m p l e t e d f o r ’{ a r t i f a c t } ’ by u s e r
’{ user Name } ’ . ” } ]
With event INFO tag start:
• The method for collecting the event is logging
journal
• The location for collecting the event is
C:\Program Files(x86)
\Siemens\WinCC\Diagnose
• The patterns within the event where contextual
information should appear is [time] INFO
[task] task started for ’artifact’ by
user ’userName’. By analyzing these event
patterns, we can collect contextual information
(time, task name, artifact name, user name),
which is important to create the Action Linkage
table and coordinate the workflow of the target
artifacts.
By analyzing historical connections, we identified
common pairs of application and PMS events. For
instance, in Table 4, the application event INFO tag
start from SIMATIC WinCC V8 — indicating the
initiation of a job identified by a specific tag — fre-
quently corresponds with the PMS event startTask in
jBPM, which starts the task in the PMS. Similarly, the
relationship between INFO tag complete and endTask
was also observed.
From Table 4, we can observe the importance
of contextual information in determining the corre-
sponding process task. The first and third lines of
Table 4 have a similar pair of application and PMS
event, INFO tag start and startTask respectively.
The generated contextual information is designed to
accurately reflect the meaning of each correspond-
ing task while avoiding ambiguity between tasks with
similar meanings. Our fine-tuned AI model can gen-
erate unique contextual information for each task.
MODELSWARD 2025 - 13th International Conference on Model-Based Software and Systems Engineering
106

Table 4: Action Linkage generation for the configuration of
Manufacture Brake system for Self-driving Car project on
SIMATIC WinCC V8 application and jBPM PMS.
App
Event
Context
Info
PMS
Event
Task
INFO
tag start
brake
system
design
start-
Task
Design brake
system blueprint
INFO tag
complete
brake
system
design
end-
Task
Design brake
system blueprint
INFO
tag start
brake
activation
start-
Task
Implement Remote
Sensing Devices
braking activation
INFO tag
complete
brake
activation
end-
Task
Implement Remote
Sensing Devices
braking activation
INFO
tag start
brake
Vehicle
start-
Task
Connect Remote
Sensing Devices
with brake control
system
INFO tag
complete
brake
Vehicle
end-
Task
Connect Remote
Sensing Devices
with brake control
system
INFO
tag start
au-
tom/brake
start-
Task
Assemble the
system into a
prototype vehicle
INFO tag
complete
au-
tom/brake
end-
Task
Assemble the
system into a
prototype vehicle
INFO
tag start
technician
Assemble
start-
Task
Assemble the
system into the
final version
INFO tag
complete
technician
Assemble
end-
Task
Assemble the
system into the
final version
Table 5: Similarity score for contextual information gener-
ated in the case study evaluation.
Contextual Info
ROUGE-
1
ROUGE-
L
BERT
score
brake system design 0.857 0.571 0.600
brake activation 0.5 0.5 0.528
brake Vehicle 0.2 0.2 0.455
autom/brake 0.0 0.0 0.393
technician
Assemble
0.222 0.222 0.462
For example, brake system design succinctly repre-
sents the task Design brake system blueprint with a
ROUGE-1 similarity score of 0.857. The generated
contextual information remains distinct even for tasks
with similar meanings, such as Assemble the system
into a prototype vehicle and Assemble the system into
the final version. The generated Action Linkage table
solves the RQ2 by successfully mapping the low-level
action in the applications and the high-level task in the
PMS, allowing seamless process monitoring.
Furthermore, users can modify the Action Link-
age table to suit their needs. For instance, if the
roles of Mechanical Engineer and Embedded System
Developer do not utilize SIMATIC WinCC V8 for
the tasks Design brake system blueprint and Imple-
ment Remote Sensing Devices braking activation re-
spectively, they can remove the related tuples from
the generated Action Linkage table. Additionally,
observed from Table 5, some generated contextual
information values, such as autom/brake and tech-
nician.Assemble, have low similarity scores on the
ROUGE-1 and ROUGE-L, but significantly better
score on BERT score. The reason for such scores
might come from the limited size of the original text.
They might be unclear or hard to remember, poten-
tially requiring human intervention to correct. The
corrected terms become learning material for pMage
to improve the Intelligent Connector suggestion and
the Intelligent Monitor training data.
The case study illustrates the capabilities of our
proposed framework, showcasing its practical bene-
fits and efficiency in solving both research questions
RQ1 and RQ2. We are currently conducting addi-
tional process monitoring studies to evaluate the ap-
plication of pMage in other domains.
7 CONCLUSION
In this study, we address the research questions of
(RQ1) how to manage and exploit the heterogeneity
of tools and PMS when establishing connections, and
(RQ2) how to deduce high-level tasks from the nu-
merous low-level actions. Through comprehensive
analyses and the application of two advancements In-
telligent Connector and Intelligent Monitor, our ex-
periments’ results demonstrate that the improved in-
termediate framework pMage not only leverages his-
torical connections to suggest configurations for new
projects using Jaccard similarity but also develops
a more efficient mechanism for generating contex-
tual information for the Action Linkage table by fine-
tuning a pre-trained AI model using domain-related
terms extracted from web pages and historical con-
nections. Although the proposed solutions have some
limitations including periodic rather than real-time
domain vocabulary updates, as collecting and enrich-
ing the vocabulary and fine-tuning the AI model for
specific tasks require considerable time, the findings
suggest that intelligent components’ integration is a
viable and more effective solution for addressing the
coordination between diverse tools and PMSs and
embracing process awareness in user workspace, ul-
timately paving the way for further exploration and
implementation in intelligent process management.
An Automated and Intelligent Interface Embracing Process Awareness into User Workspace
107

Our vision is to evolve towards a comprehensive
large process model (Kampik et al., 2023). Future
research contributions aim to use historical data for
process mining and insight discovery, train AI mod-
els to recommend new process definitions without
domain expertise and define an advanced framework
with even more assistance for the setup phases for col-
laborative teams.
REFERENCES
Baresi, L., Di Ciccio, C., Mendling, J., Meroni, G., and
Plebani, P. (2017). martifact: an artifact-driven pro-
cess monitoring platform. In Robert Claris
´
o, Hen-
rik Leopold, Jan Mendling, Wil M. P. van der Aalst,
Akhil Kumar, Brian T. Pentland, Mathias Weske,
editor, Proceedings of the BPM Demo Track and
BPM Dissertation Award co-located with 15th In-
ternational Conference on Business Process Model-
ing (BPM 2017), September 13, 2017., pages 1 – 5,
Barcelona, Spain. CEUR Workshop Proceedings.
Baresi, L., Meroni, G., and Plebani, P. (2016). A gsm-based
approach for monitoring cross-organization business
processes using smart objects. In Business Pro-
cess Management Workshops, pages 389–400, Cham.
Springer International Publishing.
Cohn, D. and Hull, R. (2009). Business artifacts: A data-
centric approach to modeling business operations and
processes. IEEE Data Eng. Bull., 32:3–9.
Delgado, A., Calegari, D., and Arrigoni, A. (2016). To-
wards a generic bpms user portal definition for the
execution of business processes. Electronic Notes
in Theoretical Computer Science, 329:39–59. CLEI
2016 - The Latin American Computing Conference.
Devlin, J., Chang, M.-W., Lee, K., and Toutanova, K.
(2019). BERT: Pre-training of deep bidirectional
transformers for language understanding. In Burstein,
J., Doran, C., and Solorio, T., editors, Proceedings
of the 2019 Conference of the North American Chap-
ter of the Association for Computational Linguistics:
Human Language Technologies, Volume 1 (Long and
Short Papers), pages 4171–4186, Minneapolis, Min-
nesota. Association for Computational Linguistics.
Due
˜
nas, S., Cosentino, V., Robles, G., and Gonzalez-
Barahona, J. M. (2018). Perceval: Software project
data at your will. In 2018 IEEE/ACM 40th Interna-
tional Conference on Software Engineering: Compan-
ion (ICSE-Companion), pages 1–4.
Dumas, M. (2011). On the convergence of data and process
engineering. In Proceedings of the 15th International
Conference on Advances in Databases and Informa-
tion Systems, ADBIS’11, page 19–26, Berlin, Heidel-
berg. Springer-Verlag.
Dumas, M., Fournier, F., Limonad, L., Marrella, A., Mon-
tali, M., Rehse, J.-R., Accorsi, R., Calvanese, D.,
De Giacomo, G., Fahland, D., Gal, A., La Rosa, M.,
V
¨
olzer, H., and Weber, I. (2023). AI-augmented busi-
ness process management systems: A research man-
ifesto. ACM Trans. Manage. Inf. Syst., 14(1):11:1–
11:19.
Evermann, J., Rehse, J.-R., and Fettke, P. (2017). Predicting
process behaviour using deep learning. Decision Sup-
port Systems, 100:129–140. Smart Business Process
Management.
Kampik, T., Warmuth, C., Rebmann, A., Agam, R., Egger,
L. N. P., Gerber, A., Hoffart, J., Kolk, J., Herzig, P.,
Decker, G., van der Aa, H., Polyvyanyy, A., Rinderle-
Ma, S., Weber, I., and Weidlich, M. (2023). Large
process models: Business process management in the
age of generative AI.
Leontjeva, A., Conforti, R., Di Francescomarino, C., Du-
mas, M., and Maggi, F. M. (2015). Complex symbolic
sequence encodings for predictive monitoring of busi-
ness processes. In Motahari-Nezhad, H. R., Recker, J.,
and Weidlich, M., editors, Business Process Manage-
ment, pages 297–313, Cham. Springer International
Publishing.
Lewis, M., Liu, Y., Goyal, N., Ghazvininejad, M., Mo-
hamed, A., Levy, O., Stoyanov, V., and Zettlemoyer,
L. (2019). BART: Denoising sequence-to-sequence
pre-training for natural language generation, transla-
tion, and comprehension.
Metzger, A., Neubauer, A., Bohn, P., and Pohl, K. (2019).
Proactive process adaptation using deep learning en-
sembles. In Giorgini, P. and Weber, B., editors, Ad-
vanced Information Systems Engineering, pages 547–
562, Cham. Springer International Publishing.
Nguyen, M. K., Nhi Tran, H., Ober, I., and Abualsaud, R.
(2024). Ai-augmented framework to enable process
awareness in collaborative teams. In 2024 Interna-
tional Joint Conference on Neural Networks (IJCNN),
pages 1–8, Yokohama, Tokyo. IEEE.
Nguyen, M. K., Tran, H. N., Abualsaud, R., and Ober, I.
(2023). An AI-enhanced process mining framework
for software process insights. In 2023 15th Interna-
tional Conference on Knowledge and Systems Engi-
neering (KSE), October 18-20, 2023, Ha Noi, Viet-
nam, pages 1–6, Hanoi, Vietnam. IEEE. ISSN: 2694-
4804.
Pennington, J., Socher, R., and Manning, C. (2014). Glove:
Global vectors for word representation. volume 14,
pages 1532–1543.
Van der Aalst, W. M. (2021). Hybrid intelligence: to au-
tomate or not to automate, that is the question. Inter-
national Journal of Information Systems and Project
Management, 9(2):5–20.
Weinzierl, S., Zilker, S., Dunzer, S., and Matzner, M.
(2024). Machine learning in business process manage-
ment: A systematic literature review. Expert Systems
with Applications, 253:124181.
MODELSWARD 2025 - 13th International Conference on Model-Based Software and Systems Engineering
108