When a model gives you mixed signals: cognitive effects and visual behavior

Abbad-Andaloussi, Amine; Franceschetti, Marco; López, Hugo A.; Schreiber, Clemens; Weber, Barbara Article — Published Version When a model gives you mixed signals: cognitive effects and visual behavior Process Science Provided in Cooperation with: Springer Nature Suggested Citation: Abbad-Andaloussi, Amine; Franceschetti, Marco; López, Hugo A.; Schreiber, Clemens; Weber, Barbara (2025) : When a model gives you mixed signals: cognitive effects and visual behavior, Process Science, ISSN 2948-2178, Springer International Publishing, Cham, Vol. 2, Iss. 1, https://doi.org/10.1007/s44311-025-00022-8 This Version is available at: https://hdl.handle.net/10419/330732 Standard-Nutzungsbedingungen: Die Dokumente auf EconStor dürfen zu eigenen wissenschaftlichen Zwecken und zum Privatgebrauch gespeichert und kopiert werden. Sie dürfen die Dokumente nicht für öffentliche oder kommerzielle Zwecke vervielfältigen, öffentlich ausstellen, öffentlich zugänglich machen, vertreiben oder anderweitig nutzen. Sofern die Verfasser die Dokumente unter Open-Content-Lizenzen (insbesondere CC-Lizenzen) zur Verfügung gestellt haben sollten, gelten abweichend von diesen Nutzungsbedingungen die in der dort genannten Lizenz gewährten Nutzungsrechte. Terms of use: Documents in EconStor may be saved and copied for your personal and scholarly purposes. You are not to copy documents for public or commercial purposes, to exhibit the documents publicly, to make them publicly available on the internet, or to distribute or otherwise use the documents in public. If the documents have been made available under an Open Content Licence (especially Creative Commons Licences), you may exercise further usage rights as specified in the indicated licence. https://creativecommons.org/licenses/by/4.0/ RESEARCH Open Access © The Author(s) 2025. Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit h t t p : / / c r e a t i v e c o m m o n s . o r g / l i c e n s e s / b y / 4 . 0 / . Abbad-Andaloussi et al. Process Science (2025) 2:19 https://doi.org/10.1007/s44311-025-00022-8 † A m i n e Abbad-Andaloussi and Marco Franceschetti contributed equally to this work. *Correspondence: Amine Abbad-Andaloussi [email protected] Marco Franceschetti marco[email protected] 1University of St.Gallen, St. Gallen, Switzerland 2Technical University of Denmark, Kgs. Lyngby, Denmark 3Karlsruhe Institute of Technology, Karlsruhe, Germany When a model gives you mixed signals: cognitive effects and visual behavior AmineAbbad-Andaloussi1*†, MarcoFranceschetti1*†, Hugo A.López2, ClemensSchreiber3 and BarbaraWeber1 Introduction Business process models are widely used to improve the understanding of business processes by combining information visualization principles with the semantics of formal modeling languages. Several process modeling languages, such as Business Process Model and Notation (BPMN)(OMG 2014) and Petri nets(Reisig 2012), offer formally defined semantics and are commonly adopted for this purpose. However, there is growing recognition that business process models, even with formal semantics, may be prone to ambiguity(Dijkman etal. 2008; Fan etal. 2016; Pittke etal. 2015), raising the need for further investigation into its impact. Ambiguity in a process model is a phenomenon that can lead to multiple valid interpretations of the process: for an example of lexical ambiguity affecting activity labels, an activity labeled Schedule review can be interpreted as an activity of scheduling a review or as an activity of reviewing a schedule. A recent study inFranceschetti etal. (2023) examined ambiguity across different artifacts within the Business Process Management (BPM) lifecycle, including informal specifications, Abstract Ambiguity in business process models can result in multiple interpretations by model readers. This leads to undesirable outcomes such as misunderstandings, unclear allocation of responsibilities, and unexpected behaviors. Despite these potential consequences, the impact of ambiguity on model readers has received limited attention so far. This article presents an eye-tracking study designed to investigate the effects of various types of ambiguity (i.e., layout, semantic, syntactic, and lexical) on readers’ cognitive load, comprehension, and visual associations while interpreting process models. In addition, the study delves into the behaviors of model readers when resolving ambiguity in process models. These behaviors are investigated following a qualitative approach combining both eye-tracking and think-aloud data. The results demonstrate that ambiguities significantly influence cognitive load, comprehension, and visual associations, emphasizing the negative effects of ambiguity. Moreover, the qualitative insights suggest that participants exhibit specific behaviors when trying to resolve ambiguities. These findings underscore the need for advanced mechanisms to detect and mitigate ambiguity in process models. Keywords Ambiguity, Process models, Eye-tracking, Cognitive load, Visual behavior Process Science Page 2 of 38Abbad-Andaloussi et al. Process Science (2025) 2:19 models, and event logs. The study highlighted the pervasive nature of ambiguity in process models through examples drawn from literature and public datasets. As process models are frequently adopted as a means of communication between stakeholders, the existence of multiple valid interpretations can undermine their communication effectiveness due to the potential mismatch between a reader’s interpretation and the modeler’s intended message. Indeed, as detailed inFranceschetti etal. (2023), ambiguity in process models entails the risk of misunderstandings, unclear responsibilities, unexpected behaviors, as well as cascading effects across other process-related artifacts managed in the BPM lifecycle. Consequently, ambiguity in process models can compromise recognized benefits of adopting Process-aware Information Systems such as clear process-based communication between stakeholders, increased efficiency, reduced redundancies, and monitoring support(Dumas etal. 2018), by diminishing these benefits(Kindler 2009; Oppl 2016). Therefore, correctly comprehending process models is imperative for the successful engineering and operation of a Process-aware Information System(Oppl 2016). Previous research has examined various types of ambiguities and strategies to reduce them in business process models at design time (cf.Fan etal. (2016); Mendling etal. (2010b); Pittke etal. (2015)). However, to the best of our knowledge, no prior study has validated the impact of ambiguity on the cognitive and behavioral aspects of process model readers. On the one hand, prior studies have explored the impact on these aspects in relation to model complexity(Petrusel etal. 2017), model quality(Heggset etal. 2015), notational deficiencies in modeling languages(Figl etal. 2013), or reader-specific attributes like modeling experience and process knowledge(Mendling etal. 2012). On the other hand, the specific impact of ambiguity on the cognitive and behavioral aspects remains, to date, unexamined. Unlike aspects such as model correctness, ambiguity is characterized by the potential for multiple equally valid interpretations. While in certain contexts, such as normative processes, this feature is intentionally incorporated to provide flexibility in model interpretation, in other contexts ambiguity can have an unintended negative impact as it generates confusion(Franceschetti etal. 2023). This study focuses on understanding how this unique characteristic, which can lead to multiple interpretations, influences the cognitive and behavioral aspects of process model readers. Particularly, in this study, we address the challenge of measuring the cognitive impact of ambiguity in process models on readers’ cognitive load (i.e., mental effort), comprehension, and visual associations (i.e., shifts of attention between model elements, which suggests an increased mental demand for integrating information, cf.Bera etal. (2019)). Additionally, we explore the visual behavior of model readers when confronted with ambiguity. To achieve this, we use eye-tracking, a method that has proven effective in previous research on process model comprehension tasks, offering valuable insights into the cognitive and behavioral aspects of model readers (cf.Bera etal. (2019); Petrusel and Mendling (2013); Petrusel etal. (2016); Wang etal. (2022); Winter etal. (2023)). Overall, we address the following research questions: •RQ1. How do different ambiguities affect model readers’ cognitive load, comprehension and visual associations? •RQ2. What patterns of visual behavior do model readers exhibit when resolving ambiguities in process models? Page 3 of 38Abbad-Andaloussi et al. Process Science (2025) 2:19 This paper extends our prior work presented inFranceschetti etal. (2024), which investigated RQ1, by additionally investigating RQ2. To answer RQ1, we investigated the eye movements of model readers while they performed different comprehension tasks on process models with and without ambiguities. To answer RQ2, we performed a qualitative study in which we observed the eye movements of model readers when confronted with ambiguity, and triangulated these observations with the verbal information provided by the model readers during retrospective think-aloud sessions. Our results demonstrate the usefulness of eye-tracking to detect ambiguity and that ambiguous process models lead to higher cognitive load, challenged comprehension, and increased visual associations (RQ1) when compared to process models without ambiguities. Moreover, our qualitative study revealed five distinct visual behavior patterns that emerge when model readers are confronted with ambiguity, each associated with varying levels of visual attention, ranging from random to highly focused (RQ2). Having demonstrated the negative effects of ambiguity and how it influences model readers’ cognitive aspects and visual behavior, future work could focus on developing methods and techniques to support model readers by automatically detecting ambiguities in process models, or assisting them in focusing their attention on model elements that provide disambiguation cues. Our results also stimulate further studies on the impact of ambiguity in relation to expertise and on disambiguation strategies. This paper is structured as follows: in Background and related worksection, we recall background concepts and formally define ambiguities in process models. In Research methodsection, we report on our study design. In Findingssection, we report on the findings. In Discussionsection, we elaborate on a discussion of the findings, the implications of the study results, and threats to validity. In Conclusionsection, we conclude the paper. Background and related work In this section, we first establish the theoretical foundations of ambiguity in BPM, formalizing four ambiguity types found in process models (cf. Ambiguity in process modelssection). We then set the theoretical foundations on cognitive theories relevant to our empirical study, setting the underpinnings for and motivating our study design (cf. Cognitive load, comprehension, and visual associationssection). Finally, we provide an overview of qualitative approaches investigating modelers’ visual behavior when engaging with process models (cf. Qualitative analysis of visual behavior during process comprehensionsection). Ambiguity in process models In general, ambiguity in BPM refers to the potential for a business process representation to yield multiple admissible interpretations. According to the characterization provided inFranceschetti etal. (2023), specifically ambiguity in process models may arise from intrinsic factors (i.e., specific to the modeling language, such as the inability to represent the resources perspective in standard Petri nets) or extrinsic factors (i.e., related to the modeling task, such as an underspecified gateway condition by an inexperienced modeler not validated by the modeling tool). As process models are formalized using modeling languages that combine both textual and graphical elements, they encompass the following aspects: syntactic aspects in Page 4 of 38Abbad-Andaloussi et al. Process Science (2025) 2:19 relation to the use of the modeling language grammar, semantic aspects in relation to the modeled behavior, lexical aspects in relation to textual elements, and layout aspects in relation to the graphical presentation. Since all these aspects influence a model reader’s interpretation of the model, in this study we aim to extend the aforementioned ambiguity characterization(Franceschetti etal. 2023) by defining layout1, semantic, syntactic, and lexical ambiguities in process models. To the best of our knowledge, these types of ambiguities are still not clearly defined in the BPM literature. To define these ambiguities in the BPM context, we considered two main options. On the one hand, we could adopt corresponding definitions from the field of linguistics(Berry And Kamsties 2004; Fortuny And Payrató 2024; Sennet 2011). However, such definitions apply specifically to natural language, which is typically presented in oral or textual form, and do not fully apply to the diagrammatic nature of process models, which combine both textual and graphical elements. Thus, definitions from linguistics do not fully capture the ways in which ambiguity arises in business process models, since process model interpretation depends not only on natural language but also on diagrammatic conventions, domain mappings, and layout (Figl 2017; Haisjackl etal. 2015). Moreover, the distinction between the different types of ambiguity in linguistics does not fully align with the one from our BPM field. For example, the BPM literature includes layout characteristics as part of pragmatic aspects(Haisjackl etal. 2015), whereas in linguistics pragmatics concerns contextual and inferential aspects of communication related to speaker intent(Fortuny And Payrató 2024). On the other hand, we could also rely on prior BPM literature that has dealt with some of these ambiguities (cf.Amna and Poels (2022); Fan etal. (2016); Haisjackl etal. (2015); Leopold etal. (2010); Mendling etal. (2010b); Pittke etal. (2015)), but has not achieved consolidated BPM-specific definitions. Moreover, prior literature that has studied model quality aspects in BPM could also be relevant (cf.Krogstie (2012, 2016); Lindland etal. (1994)) as the violation of quality guidelines often leads to ambiguity. However, ambiguity and quality are different concepts. Indeed, while ambiguities are often associated with quality issues, in general ambiguity is not reducible to quality: even a formally correct and high-quality model may give rise to ambiguity. This is observed, for example, in models of legal processes, which are of high-quality but can intentionally be designed ambiguous to allow for a flexible application of the law(Hildebrandt 2018; Slaats etal. 2013). As a result, we decided to develop our own BPM-specific definitions of ambiguity, drawing from the non-consolidated definitions in existing BPM literature, while also being informed by process model quality frameworks. Definition 1 (Layout Ambiguity) (Adapted fromAmna and Poels (2022); Haisjackl etal. (2015)) Layout ambiguity is a phenomenon that occurs at the layout level causing a process model to lack clarity in one or more process perspectives, allowing for multiple interpretations without affecting the behavior of the executable process model. In our BPM-specific view, layout ambiguity refers to ambiguity that arises not from the structure or semantics of a process model per se, but from aspects related to the presentation and layout of the model. This includes, for example, overlapping edges, poorly aligned flows, or ambiguous visual groupings–which may allow for multiple plausible 1 In our previous work(Franceschetti etal. 2024), this ambiguity is called pragmatic. Here, we use the more specific term layout to better mark the distinction from pragmatic aspects in linguistics, which are not concerned with layout. Page 5 of 38Abbad-Andaloussi et al. Process Science (2025) 2:19 interpretations of the process elements. Figure1 illustrates an example of layout ambiguity. In the example, overlapping control flow edges between the activities can be observed. This overlap allows for multiple interpretations of the precedence constraints between the activities. In the specific model fragment, it is unclear whether activity “Forge piece LDF” is followed by “Drill piece LBG” or by “Extrude piece CTG”. The potential of ambiguity arising from the presentation layout was discussed inPetre (1995), however without touching upon the impact on a reader’s cognitive and behavioral aspects. Process model comprehension in relation to layout aspects was further studied in subsequent works such asFigl (2017); Haisjackl etal. (2015); Petrusel etal. (2016). Still, the effect on comprehension specifically arising from ambiguity in the layout, i.e., the multiple possible interpretations yielded by the model layout, was not considered by these studies. Therefore, the question about how (layout) ambiguities affect model readers’ cognition remained, so far, unaddressed. Definition 2 (Semantic Ambiguity) (Adapted fromFan etal. (2016); Haisjackl etal. (2015); Krogstie (2012)) Semantic ambiguity is a linguistic phenomenon related to the usage of a modeling language that occurs as a consequence of a process model lacking validity (i.e., all statements in the model are correct and related to the process) or completeness (i.e., there is a one-to-one mapping between model constructs and domain concepts), or of differences in domain and context knowledge between model creators and readers. It allows for multiple interpretations by model readers due to the readers not being able to establish clear mappings between model constructs and domain concepts. In our BPM-specific view, semantic ambiguity refers to situations where elements of a process model can be interpreted in more than one way with respect to the (formal) process model behavior, in relation to the underlying business domain. This ambiguity may result from deficiencies in semantic quality such as lack of validity or completeness, but can also arise in formally correct models due to differences in domain knowledge or context among readers. Unlike semantic ambiguity in linguistics, which focuses on multiple meanings of words or phrases, semantic ambiguity in BPM encompasses the broader challenge of aligning model elements with a shared domain understanding. Our focus lies on the interpretive consequence: that different readers may assign different meanings to the same model fragment, regardless of its formal correctness. An example of semantic ambiguity in relation to validity is illustrated in Fig.2 (left). Here, the sequence “Approve piece LDF” – “Reject piece LDF” is presented. Simply applying common sense, a model reader would naturally assume that a piece is either approved or rejected, since these verbs express opposite actions. Therefore, in the example it is unclear whether the Fig. 1 BPMN fragment of a process with layout ambiguity Page 6 of 38Abbad-Andaloussi et al. Process Science (2025) 2:19 two activities refer to different pieces or the activities are not supposed to be both executed in the same trace. An example of semantic ambiguity in relation to completeness borrowed fromFan etal. (2016) is that of a single activity “Send request” used to model two distinct types of requests in an online auction–from the seller to initiate the auction and a buyer to join the auction. The lack of a one-to-one mapping between domain concepts (two distinct activities) and model elements (one activity) results in a loss of the activity semantics which makes it unclear which request the modeled activity refers to. Semantic ambiguity in process models was studied inDijkman etal. (2008), specifically focusing on BPMN models. Semantic issues deriving from process model quality issues were investigated inHaisjackl etal. (2015); Krogstie (2012). To reduce semantic ambiguity in process models, a possible approach is to leverage semantic information about the process domain represented through ontologies, as proposed by the authors inFan etal. (2016). In the examples, the approach helps to (i) validate the activity arrangement, detecting that “Approve piece LDF” and “Reject piece LDF” in the sequence correspond to mutually exclusive ontology classes, and (ii) identify construct excess, detecting the mismatch between one modeled activity and two separate ontology classes (cf.Fan etal. (2016)). Definition 3 (Syntactic Ambiguity) (Adapted fromAmna and Poels (2022)) Syntactic ambiguity is a grammatical phenomenon that occurs when a fragment of a process model M formalized in a modeling language L can be parsed using more than one grammatical structure of L , allowing for multiple possible interpretations of M. In our BPM-specific view, syntactic ambiguity is concerned with the parsing of the modeling language constructs. Figure2 (right) illustrates an example of syntactic ambiguity. In the figure, the control flow splits at an XOR-gateway, which has no conditions attached. Therefore, the gateway can be interpreted either as an underspecified (i.e., with unknown condition) XOR-gateway, which makes the subsequent activities mutually exclusive, or as an incorrectly assigned AND-gateway, which enforces that the subsequent activities are both executed in all traces. Syntactic ambiguity in process descriptions was investigated inAmna and Poels (2022), while the potential emergence of ambiguity due to syntactic quality deficiencies in BPMN models was investigated inHaisjackl etal. (2015). The relation between syntactic aspect of process models and model understandability was discussed inCorradini etal. (2018), along with the proposal of a set of modeling guidelines to avoid the emergence of syntactic ambiguity. Despite these studies, to the best of our knowledge, the relation between syntactic ambiguity and the cognitive and behavioral aspects of model readers remains, to date, unaddressed. Fig. 2 BPMN fragments of processes with semantic (left) and syntactic (right) ambiguities Page 7 of 38Abbad-Andaloussi et al. Process Science (2025) 2:19 Definition 4 (Lexical Ambiguity) (Adapted fromPittke etal. (2015)) Lexical ambiguity is a linguistic phenomenon related to the usage of the natural language that occurs when a textual label in a process model can be interpreted in multiple ways. According to the studies presented inCorradini etal. (2018); Pustejovsky (1998), lexical ambiguity in BPM can be caused by the use of abbreviations, homonyms, synonyms, and polysemic words (i.e, words with multiple meanings). For example, consider an activity labeled “Receive report” followed by an activity labeled “Evaluate summary”. Here, it is unclear whether the terms “report” and “summary” are used as synonyms and refer to the same data object, or they refer to different data objects. The use of specific labeling styles, such as passive voice or noun form of verbs, might as well lead to lexical ambiguity(Mendling etal. 2010b). Consider the example borrowed fromPittke etal. (2015) of an activity labeled “Plan integration”: the activity could be interpreted either as the planning of some integration, or as the integration of some plan. Prior work from Mendling et al. investigated the usage of activity labels and the resulting lexical ambiguity(Mendling etal. 2010a, b). These studies resulted in guidelines(Mendling etal. 2010c), refactoring recommendations(Leopold etal. 2010), and automatic approaches to detect and resolve lexical ambiguity in process models(Pittke etal. 2015). To the best of our knowledge, however, these studies did not explicitly investigate the impact on the specific cognitive and behavioral aspects that we set to measure in this paper. Cognitive load, comprehension, and visual associations In this section, we provide background and set the theoretical underpinnings to investigate the impact of different ambiguities in process models on readers’ cognitive load, comprehension, and visual associations. This in turn will allow us to address RQ1 (cf. Introductionsection). Cognitive Load. Cognitive load denotes the workload imposed on the human working memory during tasks requiring mental processing(Sweller 2011). The Cognitive Load Theory (CLT) discerns three types of cognitive load: Intrinsic, Extraneous, and Germane(Sweller 2011; Chen etal. 2016). Intrinsic load emerges from the essential complexity of the process model, which is inherent to the encoded process specifications. Extraneous load, in turn, emerges from the accidental complexity of the process model, which is typically associated with the way the model is presented to the user. Finally, Germane load arises from the difficulty to integrate the information extracted from the model with ones’ mental schema in order to develop an overarching understanding of the process. Typically, readers attempt to identify, in the material at hand, features that signal association with their pre-existing mental schemas (An 2013). This identification facilitates schema activation, defined as the process through which individuals retrieve and apply previously acquired knowledge structures from memory to interpret new information (An 2013). For instance, when reading a process model, they may associate an XOR gateway with their pre-existing schema related to mutual exclusion. However, when ambiguities are present, readers may struggle to identify and activate the appropriate schema because multiple interpretations become plausible. As depicted in Fig. 2 (right), an XOR gateway with unlabeled outgoing edges creates ambiguity. Readers might either rely on contextual information to select the most suitable branch, hence associating the underspecified XOR gateway with their mutual-exclusion schema. Alternatively, they Page 8 of 38Abbad-Andaloussi et al. Process Science (2025) 2:19 can interpret the XOR gateway as an AND gateway, where unlabeled outgoing edges make sense, hence associating the gateway with their pre-existing schema on concurrency. This suggests that ambiguities can challenge the integration of new information with existing schemas, consequently increasing germane cognitive load. This proposition is supported by Campbell’s work(Campbell 1988). In his literature review, Campbell identified several characteristics of complex tasks that impose high mental demands (i.e., cognitive load) on readers. One such characteristic is the presence of uncertain or conflicting information within an artifact. Another is the availability of multiple viable solutions to the task. This, in turn, increases readers’ cognitive load as they must look for additional information, assess alternatives, and make decisions among similarly viable options. Existing literature proposed several measures aimed at estimating readers’ cognitive load(Figl 2017; Figl etal. 2024; Figl And Laue 2015; Abbad-Andaloussi etal. 2023a; Wang etal. 2022; Schreiber etal. 2024; Zugal 2013; Weber etal. 2021). Among them, self-reported measures rely on individuals’ own assessment of perceived difficulty(Figl 2017; Figl And Laue 2015; Abbad-Andaloussi etal. 2023a; Wang etal. 2022; Schreiber etal. 2024; Zugal 2013), which for example can be rated by readers using a 5-point Likert scale (ranging from 0: “very easy” to 4: “very difficult”) after completing a task(Schreiber etal. 2024; Zugal 2013). Beside self-assessment, which can be subjective, eye-tracking measures can provide reliable insights into cognitive load(Holmqvist etal. 2011; Figl etal. 2024; Abbad-Andaloussi etal. 2023a; Schreiber etal. 2024). Eye-tracking enables the analysis of fixation characteristics of readers (i.e., the amount of time the eye remains stationary at a specific position of the stimulus, e.g., a process model(Holmqvist etal. 2011)) to estimate their cognitive load(Holmqvist etal. 2011). The use of fixation features as indicators of cognitive load is grounded in the eye-mind hypothesis(Holmqvist etal. 2011), which postulates that the mind processes the content currently fixated by the eyes. Glöckner and Herbold (2008) extended this theory by suggesting that fixations lasting ≥250ms signify mental processing, which can be linked to cognitive load. While the eye–mind hypothesis was challenged, particularly because of its limitations in accounting for asynchronies between attention and eye movements, as well as its inability to capture the influence of peripheral vision on cognitive processing(Holmqvist etal. 2011), it continues to serve as a central assumption in research exploring how users engage with software artifacts. This is evident in a wide range of studies that adopt the hypothesis to examine various cognitive and behavioral aspects of users’ interactions with software artifacts (overview inSharafi etal. (2015); Weber etal. (2021); BatistaDuarte etal. (2021)). To evaluate the effect of ambiguities on cognitive load, we adopt both the self-assessment of perceived difficulty measure and the eye-tracking measures reflecting cognitive load. For the former, we use a 5-point Likert scale questionnaire to capture perceived difficulty. For the latter, we use the mean number of fixations lasting ≥250ms as an indicator of cognitive load. We hypothesize that both measures will exhibit significant increases when readers are confronted with ambiguous process models, reflecting the additional mental effort required to resolve the ambiguities. Comprehension. In the cognitive science literature, very high levels of cognitive load can impair readers’ performance, particularly in terms of task accuracy (Veltman And Jansen 2005) and response time (Chen etal. 2016). Moving to the process model Page 15 of 38Abbad-Andaloussi et al. Process Science (2025) 2:19 task allowed participants to get accustomed with the interface and task format without influencing the study results, as the outcomes of this task were excluded from the subsequent data analysis. To address potential learning and fatigue effects, the presentation order of the remaining tasks was randomized. This randomization ensured that learning effects and participant fatigue were distributed evenly across the different experiment tasks. After completing each task, participants rated the perceived difficulty of solving the task for each sub-process using a Likert scale. These ratings were used to compute the perceived difficulty measure of cognitive load (cf. Study designsection). Participants were also asked retrospectively to state whether they encountered any issues while solving the task and to describe how they overcame them. This was used to inform us whether they had noticed the ambiguities in the process models and how they resolved them. We used this information in the data analysis to consider only tasks where ambiguity was noticed and to qualitatively analyze participants’ behavior when resolving ambiguities (cf. Data collection and analysissection). Data collection and analysis The data was gathered using EyeMind(Abbad-Andaloussi etal. 2023b), which is an eye-tracking tool designed for capturing eye-tracking data on process models displayed within an interactive editor. This tool allows users to seamlessly navigate various parts of a model and explore its sub-processes. A significant advantage of EyeMind is its ability to support dynamic eye-tracking stimuli, enabling users to freely browse through different views, scroll, and zoom in various parts of the stimulus(Abbad-Andaloussi etal. 2023b; Holmqvist etal. 2011). Conducting experiments with dynamic stimuli is recognized as a complex and time-intensive process(Abbad-Andaloussi etal. 2023b; Holmqvist etal. 2011). Consequently, researchers often rely on static stimuli, using small, non-interactive process models presented as images. While simpler to implement, this approach does not accurately represent the complexity and usability of real-world process models, limiting the ecological validity (i.e., ability to generalize findings)(Abbad-Andaloussi etal. 2023b; Holmqvist etal. 2011). To overcome the limitations of relying on static stimuli and better reflect the size and complexity of real-world models, we selected EyeMind. To the best of our knowledge, it is the only tool capable of providing this functionality for process models. After data collection, we selected the trials3 in which participants correctly identified the ambiguities embedded in the models. This selection was based on the assumption that non-identified ambiguities would have no effect on participants’ cognitive and behavioral responses. Furthermore, this selection was based on our goal to investigate the effects of recognized ambiguity as the enabler of the cognitive and behavioral aspects that we set to study. This is because we could not measure the effects of what was not perceived as ambiguous and did not result in such effects. Ambiguities were identified by the participants in 342 trials out of 528 trials (derived from 44 participants completing 12 tasks, excluding the familiarization task). To determine these trials, participants who stated having encountered issues in solving a task were asked to navigate through the sub-process models to show us the specific fragment responsible for their difficulties and explain us the respective reasons. We selected the trials in which the indicated fragment 3 A trial refers to an instance of a participant performing a task Page 16 of 38Abbad-Andaloussi et al. Process Science (2025) 2:19 matches with the ambiguity and the explanation could be associated with the ambiguity. In a limited number of cases, participants noted encountering additional issues that were not related to ambiguous fragments, such as finding an xor-gateway condition not immediately clear at first sight. In these cases, there were no unintended ambiguities discovered by the participants, but only issues beyond ambiguity. Since these issues were not related to ambiguity and our study focuses on the effects of ambiguity, we did not include the respective data in our analysis. To address RQ1 (cf. Introductionsection), we calculated the measures corresponding to the constructs on the dependent variables side of the research model illustrated in Fig.3. We conducted these calculations at the level of each sub-process. Since participants were assigned three tasks per ambiguity type (cf. Study designsection), three data points were collected per factor level and participant. To mitigate inter-dependencies among data points, we computed the mean value of each measure at each factor level for each participant. We computed both descriptive and inferential statistics, as reported in Table3. Descriptive statistics facilitated pairwise within-subject comparisons between the mean values of each measure across the factor levels: ambiguity and no ambiguity for each ambiguity type. To determine the statistical significance of the observed differences, we used the Wilcoxon Signed-Rank inferential Table 3 Descriptive and inferential statistics A. H. Measure Descriptive Inferential No Ambiguity Ambiguity p-value Mean Mean Layout A. H1 Cognitive Load Perceived Difficulty 1.038 3.331 < .001 Fixations >= 250 ms 32.754 74.508 <.001 H2 Comprehension Comprehension Efficiency 27198.497 51667.651 < .001 H3 Visual Associations AOI Run Count 21.159 46.357 < .001 Semantic A. H1 Cognitive Load Perceived Difficulty 0.893 2.240 < .001 Fixations >= 250 ms 28.354 41.547 < .001 H2 Comprehension Comprehension Efficiency 25809.034 39514.319 < .001 H3 Visual Associations AOI Run Count 20.067 34.573 < .001 Syntactic A. H1 Cognitive Load Perceived Difficulty 0.973 2.217 < .001 Fixations >= 250 ms 25.212 34.308 < .001 H2 Comprehension Comprehension Efficiency 26414.001 32977.135 < .001 H3 Visual Associations AOI Run Count 19.151 27.489 < .001 Lexical A. H1 Cognitive Load Perceived Difficulty 0.786 1.976 < .001 Fixations >= 250 ms 21.610 21.463 0.634 H2 Comprehension Comprehension Efficiency 16258.680 19436.565 0.019 H3 Visual Associations AOI Run Count 18.086 22.610 0.008 Comprehension Efficiency unit: milliseconds. Note: p< 0.05 informs that the pairwise difference of means between the no ambiguity and ambiguity levels is significant Page 17 of 38Abbad-Andaloussi et al. Process Science (2025) 2:19 test(Wilcoxon 1945), which is suitable for pairwise within-subject comparisons and does not impose assumptions on the normal distribution of the data. Besides validating the impact of ambiguity, we conducted a qualitative exploratory analysis, using the participants’ eye tracking and think-aloud data (cf. overview in Fig.5). Herein, the aim was to investigate their behavior when resolving ambiguity (RQ2, cf. Introductionsection). Specifically, we used a qualitative coding approach from grounded theory (Charmaz 2006; Bryant and Charmaz 2019) to identify common behaviors adopted by the participants. The codes describing visual behavior patterns were developed based on preliminary observations of the video recordings (of approximately 40 minutes each) of a sample of 23% of the study participants. An initial coding(Bryant and Charmaz 2019) phase was conducted on these videos to identify emerging visual behaviors. Herein, we assigned descriptive labels to video segments, reflecting the participants’ behavioral actions (e.g., fixating on a specific model element for a long period of time or shifting rapidly across the model), without making assumptions about the meaning of these actions. This phase allowed us to remain open to all possible interpretations and to capture the full range of behaviors participants exhibited when engaging with ambiguous process models. Afterward, we generated AOIs-order over time plots of all the trials in which ambiguities were noticed. Figure6 depicts an example of an AOIs-order over time plot. This plot visualizes how a participant’s attention shifts across different elements (represented as AOIs) of a process model while performing a task. The bars on the X-axis show the AOIs visited by the participant, sorted by their order of occurrence in time. This temporal perspective reflects the sequence of process model elements that the participant visited during the task. The black AOIs refer to the ambiguous elements of the process model, while the remaining AOIs refer to the non-ambiguous elements of the model. This latter set of AOIs have different colors, allowing, in turn, to identify those that were visited several times by the participant. This color coding helps visualize how the participant shifted Fig. 5 Overview of the qualitative exploratory analysis Page 18 of 38Abbad-Andaloussi et al. Process Science (2025) 2:19 Fig. 6 Example of a simplified AOIs-order over time plot Page 19 of 38Abbad-Andaloussi et al. Process Science (2025) 2:19 Fig. 7 Process maps comparing the visual associations of a participant (SP7) when reading a sub-process without (left) and with (right) a semantic ambiguity. A higher resolution of this figure is available in the online appendix. The circle with a dot inside denotes the process start, the double circle with a square inside denotes the process end. Rectangles refer to visits to the different process model activities; edges refer to the transitions for visiting one activity from another. The color scale of the rectangles refers to the absolute visit frequency to an activity; the thickness and labels on the edges refer to the absolute transition frequency, resp. the number of transitions between each pair of activities Page 20 of 38Abbad-Andaloussi et al. Process Science (2025) 2:19 between ambiguous and non-ambiguous elements of the model. The X-axis is additionally decorated with horizontal blocks, colored in blue, red, or green to infer respectively whether the visited AOI belongs to the first, second, or third sub-process of our model. As mentioned in Study designsection, the second sub-process was the one incorporating an ambiguity. On the Y-axis, the height of the bars defines the duration of each AOI visit (in milliseconds). Moreover, we have set 250ms as a threshold on the Y-axis, allowing to distinguish the visits reflecting mental processing (Glöckner And Herbold 2008). Projecting visit duration onto the Y-axis facilitates the comparison of how long participants engaged with ambiguous versus non-ambiguous model elements, and helps to identify visits likely associated with mental processing. All in all, the AOIs-order over time plot enables the analysis of participants’ visual behavior by revealing the sequence in which process model elements are visited (via the X-axis), the duration of each visit to a specific element (via the Y-axis), and which visits are likely to involve mental processing (Figs. 7 and 8). Following the generation of the AOIs-order over time plots, we conducted focused coding(Bryant and Charmaz 2019) on all these plots through the observation and coding of participants’ behavioral patterns, which were aligned with those observed in the videos. In this phase, we systematically categorized recurring behaviors by grouping similar initial codes into more abstract and conceptually meaningful codes. For instance, shifting short visits all over the process model elements without focusing on specific elements was coded as Sweep, while shifting long visits across a large area of the model covering several process elements was coded as Explore. When long visits occurred between a limited number of specific process elements, we assigned the code Target. Back-and-forth transitions between the ambiguity and another specific process element Fig. 8 Excerpts from ambiguous process models showing the ambiguous model fragments. The complete models are available in the online appendix Page 21 of 38Abbad-Andaloussi et al. Process Science (2025) 2:19 Fig. 9 AOIs-order over time plot showing the Sweep visual behavior of participant SP7 while trying to resolve an ambiguity, visible in the sequence of short visits ( <250ms ) to multiple process elements shown in different colors between the first and the last visits to the ambiguity (shown in black) Page 22 of 38Abbad-Andaloussi et al. Process Science (2025) 2:19 were captured under the code Bounce, whereas mostly uninterrupted sequences of long visits to the ambiguous process elements themselves were labeled Hover. These focused codes allowed us to infer the common strategies participants used when attempting to resolve ambiguities. To ensure the robustness of our codes, the coding was done by one author and subsequently reviewed by another author; in case of disagreements, the authors discussed the code until convergence. This coding strategy (aligned with prior literature, e.g.,Jayaraman etal. (2024)) was intentionally chosen over independent coding considering the inherent complexity of the AOIs-order over time plots, whose visual analysis requires collaborative effort, iterative back-and-forth discussions, and careful inspection with multiple pairs of eyes to discern the behavioral patterns (cf. Figs.9, 10, 11, 12, 13 and 14). Moreover, to ensure the validity of our patterns identification, we incorporated as supplementary evidence the verbal data obtained by asking retrospectively at the end of each task which strategies the participants employed to overcome any difficulties encountered. To do this, we listened to the audio track of the video recordings of the data collection sessions, timestamping the periods where relevant answers were provided, and taking note of the answers. This allowed us to triangulate the coded behaviors with the verbal insights provided by the participants. Finally, we applied axial coding to establish relationships between our codes(Bryant and Charmaz 2019). In doing so, we organized the focused codes into a scale of attention focus, arranging the identified behaviors according to their associated levels of attention, ranging from random attention (e.g., Sweep) to focused attention (e.g., Hover). This last coding phase was conducted collaboratively between the authors. Data availability and reproducibility To ensure transparency, reproducibility, and replicability, we provide an online appendix, which includes: •the full set of process models used in the tasks of the study; •a comprehensive report detailing the complexity metrics of the process models; •list of the experiment tasks with the specific guideline violations and the ambiguities these violations introduced; •demographic information of the participants; •a higher-resolution version of Figs.4, 7 and 8; •process maps illustrating the participants’ visual associations; •all AOIs-order over time plots derived from the eye-tracking data; •the qualitative codes with the behaviors observed; •the Python notebook used for data analysis, along with the corresponding results. The online appendix can be accessed at https://doi.org/10.5281/zenodo.16738478. Findings In this section, we present our findings organized by research question. RQ1. How do different ambiguities affect model readers’ cognitive load, comprehension and visual associations? Cognitive load. Cognitive load was measured using perceived difficulty and mean number of fixations with duration ≥250ms , collected at Page 23 of 38Abbad-Andaloussi et al. Process Science (2025) 2:19 Fig. 10 AOIs-order over time plot showing the Explore visual behavior of participant SP6 while trying to resolve an ambiguity, visible in the sequence of long visits ( ≥250ms ) to multiple process elements shown in different colors between the first and the last visits to the ambiguity (shown in black) Page 24 of 38Abbad-Andaloussi et al. Process Science (2025) 2:19 Fig. 11 AOIs-order over time plot showing the Target visual behavior of participant KP23 while trying to resolve an ambiguity, visible in the sequence of long visits ( ≥250ms ) to the few process elements shown in yellow, green and violet between the first and the last visits to the ambiguity (shown in black) Page 31 of 38Abbad-Andaloussi et al. Process Science (2025) 2:19 SP3, but then shifting to a Target behavior (focused on the connected activities) helped to figure some cues to make an assumption on the interpretation of the ambiguous model fragment that led to the completion of the task. The analysis of the entire dataset suggests that model readers do not always employ a single strategy, but apply multiple strategies while attempting to resolve ambiguity. Specifically, we observed Sweep behavior in combination with other behaviors in 79 trials, Explore behavior in combination with other behaviors in 60 trials, Target behavior in combination with other behaviors in 154 trials, Bounce behavior in combination with other behaviors in 42 trials, and Hover behavior in combination with other behaviors in 122 trials. Besides, in 4 out of our 342 trials, the participants’ behavior was unclear in the plots due to quality issues in the recorded data. Moreover, for any given ambiguity, we did not consistently observe the same visual behavior pattern across the participants of our study. We also did not observe a consistent adoption of the same pattern across the ambiguities for any given participant. Discussion Testing the effects of ambiguity Our findings with respect to RQ1 (cf. Introductionsection), reveal that layout, semantic, and syntactic ambiguities, which affect the process control flow, exert a significant impact on cognitive aspects. These ambiguities increase model readers’ cognitive load, reduce their comprehension, and result in a significant amount of visual associations, indicating heightened cognitive integration effort(Bera etal. 2019). In contrast, lexical ambiguities, which affect the labels, exert a less pronounced effect. While we could observe effects in terms of lower comprehension and higher visual associations, no clear effects could be observed in terms of cognitive load. Specifically, the mean number of fixations with duration ≥250ms , which are associated with mental processing (cf. Cognitive load, comprehension, and visual associationssection), did not significantly differ between models with and without lexical ambiguities. A possible explanation of this fact is that ambiguities in the lexicon may not impose as much cognitive load as ambiguities in the control flow. Exploring ambiguity resolving behaviors Our findings with respect to RQ2 (cf. Introductionsection) suggest that readers exhibit different behaviors when resolving ambiguities. Specifically, we observed and defined for the first time five distinct patterns of visual behavior induced by ambiguity in process models, namely Sweep, Explore, Target, Bounce, and Hover. These behaviors suggest different strategies of attempting to integrate information to resolve ambiguities. Building upon the findings presented in Findings section and applying axial coding(Bryant and Charmaz 2019) (cf. Data collection and analysissection and the qualitative exploratory analysis process illustrated in Fig.5), we examined the relationships between the identified visual behavior patterns and the levels of attention exhibited by the participants when resolving ambiguity. Herein, we suggest that the observed ambiguity-resolving behaviors are associated with varying degrees of attention focus. Figure15 illustrates these degrees through a continuous scale of attention focus. In the scale, Sweep reflects the least focused and most random attention, involving unfocused visits spanning the whole process model. Explore denotes a dispersed attention, as it is Page 32 of 38Abbad-Andaloussi et al. Process Science (2025) 2:19 Fig. 15 Visual behavior codes in a scale of attention. Hover suggests the most focused attention when facing ambiguity, while Sweep suggests the most random attention Page 33 of 38Abbad-Andaloussi et al. Process Science (2025) 2:19 characterized by a non-targeted exploration of multiple process model elements, suggesting that the readers’ attention is not focused toward specific process elements but toward general contextual information. Target is associated with an intermediate degree of attention, as model readers shift between a limited number of process elements, suggesting that the readers’ attention is focused toward identifying elements that can provide disambiguation cues from a selected pool of candidates. Bounce indicates a higher degree of focused attention, as model readers alternate between the ambiguous elements and a specific process element likely identified as relevant, suggesting that the readers’ attention is focused toward applying the information extracted from this element to resolve the ambiguity. Finally, Hover is associated with the highest level of focused attention, as model readers concentrate intensely on the ambiguous process elements while trying to resolve the ambiguity through a detailed element examination. As reported in Findingssection, we did not observe consistent associations between specific ambiguity types, visual behavior patterns, or participants. This suggests that the adoption of an ambiguity resolution strategy is rather subjective and varies between the concrete ambiguities a model reader is confronted with. Supporting model readers Given the demonstrated negative effects of ambiguities, it becomes crucial to explore ways to better support model readers–either by minimizing ambiguities in process models or by enhancing the readers’ ability to handle them. For instance, automated techniques for label refactoring (cf.Pittke etal. (2015)) could help to lessen the presence of lexical ambiguity. However, it is unrealistic to assume that ambiguities can be completely eliminated. As a matter of fact, certain ambiguities are intentional, for example designed to enable flexible interpretation and execution of processes (cf.Franceschetti etal. (2023)). Despite this, the significant impact of ambiguities on task performance underscores the need for support beyond merely detecting modeling errors. One potential approach is fostering a feedback loop in which modelers and model readers collaborate to identify ambiguous process elements. Alternatively, process model analysis tools could be developed to automatically detect ambiguities. While not all ambiguities might be automatically detected due to their inherently subjective nature, one can envision tools for the automatic detection of certain syntactic ambiguities based on formal definitions, such as those presented in Ambiguity in process modelssection or inFranceschetti etal. (2023), or lexical ambiguities based on Natural Language Processing. These tools might leverage ontology annotations, similar to the techniques proposed inFan etal. (2016). Another possibility is that, with further development, the cognitive effects and behavior patterns associated with the detection and resolution of ambiguity could be leveraged to develop a new generation of context adaptive systems that deploy pre-trained machine learning models (e.g.,Abbad-Andaloussi etal. (2024)) to detect when users are facing ambiguities and guide them in disambiguating the model, helping them, in turn, to move from random attention to focused one. This can, for example, be achieved by highlighting contextual information or displaying additional artifacts such as guiding annotations or simulations. Specifically, to detect ambiguities, the machine learning models can be trained using a supervised learning approach (Aggarwal 2015). Features derived from eye-tracking fixations, saccades, visits to AOIs (Holmqvist etal. 2011), Page 34 of 38Abbad-Andaloussi et al. Process Science (2025) 2:19 and other cognitive and behavioral patterns associated with ambiguity can be collected within defined time windows and paired with labels indicating whether users are experiencing ambiguity. These labels may be obtained from think-aloud protocols conducted concurrently with the task. The models can be then trained to map the extracted features to the corresponding ambiguity label at the level of each time window. At run-time, incoming data can be buffered into time windows, each fed into the trained machine learning model to predict the presence of ambiguity. If an ambiguity is detected, then contextual support can be provided by highlighting relevant information or presenting additional artifacts, such as guiding annotations or interactive simulations. Implications We identify implications for research, practice, and education underscored by our findings. Implications for Research. Regarding research, future empirical studies examining human and cognitive aspects in process modeling should account for the presence of ambiguity in process models. Ambiguities could unintentionally be a confounding factor influencing research outcomes, making it essential to mitigate them in the design of future experiments. With the AOIs-order over time plots we introduced a visualization that emphasizes the visual behaviors adopted by model readers when confronted with ambiguity. This paper presents a first attempt at using these plots toward this goal, which resulted in the identification of key patterns describing how ambiguities are resolved. By capturing two essential aspects of process model comprehension, i.e., attention to different model parts and transitions between them, these plots offer a valuable tool for future studies. Specifically, they can support deeper exploration of users’ attention distribution (through the height of the bars referring to the time spent during each visit to a process model element) and cognitive integration processes (through the order of the bars referring to their order of occurrence over time), both critical for understanding process models(Bera etal. 2019). Implications for Practice. Regarding practice, our results highlight the importance of minimizing ambiguities in process models as well as providing disambiguation cues to support model readers. This could be achieved by enriching process models with supplementary information (cf.Abbad-Andaloussi etal. (2021)) or making contextual cues in the models more explicit, since our qualitative study suggests that model readers tend to look for disambiguating cues in process models when they are confronted with ambiguity. Implications for Education. With regards to education, our results underscore the importance of raising awareness among process modeling trainees about the potential ambiguities in their models and the impact these ambiguities can have. Training programs should incorporate discussions on identifying and managing ambiguities, empowering trainees with the skills to create process models that can easily be interpreted. Threats to validity We recognize the existence of potential threats to the validity of our study. First, internal validity may be threatened by the presence of confounding factors that cannot be entirely eliminated. However, we mitigated this threat by designing a controlled experiment based on a pre-defined research model (cf. Study designsection), carefully and Page 35 of 38Abbad-Andaloussi et al. Process Science (2025) 2:19 systematically preparing the experiment materials (cf. Study designsection), randomizing the order in which the tasks were presented to avoid learning and fatigue effects (cf. Experiment proceduresection), and adhering to a strict and uniformly applied data collection protocol during all sessions of data collection (cf. Experiment proceduresection). A second threat to internal validity arises from retrospectively asking participants to state whether they encountered any issues while performing each task after its completion. Each task was based on a process model that was deliberately designed to include an ambiguity that hindered its execution. Therefore, it is possible that the participants might have anticipated the presence of issues in subsequent tasks due to this repeated questioning. Such anticipation represents a potential confounding factor, as it could influence their cognitive processing. To mitigate this threat, we took several measures. We carefully avoided disclosing the goals of our study to the participants, and avoided mentioning that they would encounter issues while analyzing the process models. We refrained from making the participants aware of the concept of ambiguity in process models, never mentioning ambiguity. Additionally, we deliberately phrased the question in a generic manner to avoid suggesting that participants were being asked to analyze process models affected by ambiguity (“Name any issues you encountered while answering the question, along with how you overcame them”). Finally, by randomizing the presentation of the tasks (cf. Study designsection) we ensured that the increased cognitive load resulting from the potential anticipation of issues was evenly distributed across all tasks. Nevertheless, in future studies, this threat could be further mitigated with the inclusion of tasks where all the models are free from ambiguity and by phrasing questions excluding nudging terms such as issues (e.g., “Name any difficulties you encountered while answering the question, along with how you overcame them”), thereby minimizing the likelihood of participants anticipating and actively seeking out problems. With regards to our qualitative analysis, a potential threat to internal validity derives from the possibility that the visual behaviors do not accurately capture the approaches employed by the participants to resolve ambiguity. To mitigate this threat, we triangulated the visual behavior patterns with the retrospective think-aloud data of the study participants when asked about the strategies adopted to overcome the difficulties in solving the given tasks. Another threat stems from potential bias in our subjective interpretation of the visual behavior patterns, represented through the AOIs-order over time plots, and the insights, extracted from the think-aloud data. To mitigate this threat, we reviewed the coding in order to validate our interpretations (cf. Data collection and analysissection). Moreover, we recognize the limitation of measuring only one of the two components of comprehension efficiency, namely the response time, neglecting comprehension accuracy. As explained in Cognitive load, comprehension, and visual associationssection, this is due to the impossibility to measure comprehension accuracy due to the inherent admissibility of the multiple interpretations of ambiguities. The external validity of our study may be threatened by the inability to generalize the experiment results due to the sample size of the participants or the modeling language used in the process models. To mitigate these threats, we recruited 44 participants, which, to the best of our knowledge, positions our study among the most extensive eyetracking studies conducted in the context of process modeling. Regarding the specific use of BPMN, we argue that the investigated ambiguities are not exclusive to BPMN. Page 36 of 38Abbad-Andaloussi et al. Process Science (2025) 2:19 Such ambiguities could also be encountered in other imperative modeling languages, such as workflow nets and EPC models(Keller etal. 1992). Conclusion Ambiguities in a process model result in multiple potential interpretations of the process. By leveraging eye-tracking, we explored how these ambiguities influence model readers during various model comprehension tasks. Our findings reveal a significant influence on cognitive load, comprehension, and visual associations. Additionally, our results indicate the adoption of specific visual behavior patterns when solving ambiguity (Sweep, Explore, Target, Bounce, and Hover), which can be associated with different levels of attention ranging from random to focused attention. The negative effects of ambiguity highlight the importance of providing adequate training and exercising caution to minimize ambiguities in process models. Additionally, they underscore the need for designing novel automated tools to assist model readers in identifying ambiguities and potentially resolving them. In future work, it is worthwhile to investigate the role played by BPM expertise in the detection of ambiguities, including in relation to diverse industrial backgrounds. Furthermore, it is also worthwhile to specifically investigate the cognitive and behavioral impact of intentional ambiguities. Moreover, given the richness of our eye-tracking data, new machine learning models can be developed to automatically detect when readers are challenged with ambiguity, and guide them resolving it. These models will provide the foundation for a new generation of ambiguity-aware context-adaptive systems. Acknowledgements We gratefully thank John Krogstie for his valuable input and insightful discussions. Authors’ contributions A.A.-A., M.F., H.A.L., and B.W. contributed to the work’s conception. A.A.-A., M.F., and C.S. conducted the experiment. A.A.-A. and M.F. analyzed the experiment results and wrote the manuscript. All authors contributed to reviewing the manuscript. Funding This work has received funding from the Swiss National Science Foundation under Grant No. IZSTZ0_208497 (ProAmbitIon project). Amine Abbad-Andaloussi is supported by the International Postdoctoral Fellowship Grant under Grant No. 1031574 from the University of St.Gallen, Switzerland. Hugo A. López is supported by research grant “Center for Digital CompliancE (DICE)” (Grant No. VIL57420) from VILLUM FONDEN. Data availability An online appendix containing supplementary material as detailed in Data availability and reproducibilitysection is available at https://doi.org/10.5281/zenodo.16738478. Declarations Competing interests The authors declare no competing interests. Consent to participate The study participants provided their informed consent to contribute with their responses and to have their eye-tracking and video/audio data recorded. Participation was voluntary, and all participants were informed of their right to withdraw at any time without penalty. The collected data was anonymized to ensure the privacy of the participants and used solely for the purposes of this study. Received: 30 December 2024 / Accepted: 25 July 2025 References Abbad-Andaloussi A, Burattin A, Slaats T et al (2023a) Complexity in declarative process models: metrics and multi-modal assessment of cognitive load. Expert Syst Appl 233:120924 Page 37 of 38Abbad-Andaloussi et al. Process Science (2025) 2:19 Abbad-Andaloussi A, Lübke D, Weber B (2023b) Conducting eye-tracking studies on large and interactive process models using eyemind. SoftwareX 24:101564 Abbad-Andaloussi A, Schreiber C, Weber B (2024) Using eye-tracking to detect search and inference during process model comprehension. In: Proceedings of the 30th International Conference on Cooperative Information Systems (in press) Abbad-Andaloussi A, Zerbato F, Burattin A et al (2021) Exploring how users engage with hybrid process artifacts based on declarative process models: a behavioral analysis based on eye-tracking and think-aloud. Softw Syst Model 20:1437–1464 Aggarwal CC (2015) Data mining: the textbook. Springer Amna AR, Poels G (2022) Ambiguity in user stories: a systematic literature review. Inf Softw Technol 145:106824 An S (2013) Schema theory in reading. Theory Pract Lang Stud (TPLS) 3(1) Batista Duarte R, Silva da Silveira D, de Albuquerque BV et al (2021) A systematic literature review on the usage of eye-tracking in understanding process models. Bus Process Manag J 27(1):346–367 Bera P, Soffer P, Parsons J (2019) Using eye tracking to expose cognitive processes in understanding conceptual models. MIS Q 43(4):1105–1126 Berry DM, Kamsties E (2004) Ambiguity in requirements specification. In: Perspectives on software requirements. Springer, pp 7–44 Bryant A, Charmaz K (2019) The SAGE Handbook of Current Developments in Grounded Theory. SAGE Publications Caldiera VRBG, Rombach HD (1994) The goal question metric approach. Encycl Softw Eng 528–532 Campbell DJ (1988) Task complexity: a review and analysis. Acad Manag Rev 13(1):40–52 Charmaz K (2006) Constructing grounded theory: A practical guide through qualitative analysis. Sage Chen F, Zhou J, Wang Y et al (2016) Robust multimodal cognitive load measurement. Springer Corradini F, Ferrari A, Fornari F et al (2018) A guidelines framework for understandable bpmn models. Data Knowl Eng 113:129–154 Dijkman RM, Dumas M, Ouyang C (2008) Semantics and analysis of business process models in bpmn. Inf Softw Technol 50(12):1281–1294 Djurica D, Kummer TF, Mendling J etal (2023) Investigating the impact of representation features on decision model comprehension. Eur J Inf Syst 1–25 Dumas M, LaRosa M, Mendling J etal (2018) Process-aware information systems. Fundam Bus Process Manag 341–369 Fan S, Hua Z, Storey VC et al (2016) A process ontology based approach to easing semantic ambiguity in business process modeling. Data Knowl Eng 102:57–77 Figl K (2017) Comprehension of procedural visual business process models. Bus Inf Syst Eng 59(1):41–67 Figl K, Laue R (2015) Influence factors for local comprehensibility of process models. Int J Hum-Comput Stud 82:96–110 Figl K, Mendling J, Strembeck M (2013) The influence of notational deficiencies on process model comprehension. J Assoc Inf Syst 14(6):1 Figl K, Soffer P, Weber B (2024) Guiding attention in flow-based conceptual models through consistent flow and pattern visibility. Decis Support Syst 185:114292 Fındık-Coşkunçay D, Çakır MP (2022) An investigation of the relationship between joint visual attention and product quality in collaborative business process modeling: a dual eye-tracking study. Softw Syst Model 21(6):2429–2460 Fortuny J, Payrató L (2024) Ambiguity in linguistics 1. Studia Linguistica 78(1):1–7 Franceschetti M, Abbad-Andaloussi A, Schreiber C etal (2024) Exploring the cognitive effects of ambiguity in process models. In: International Conference on Business Process Management. Springer, pp 493–510 Franceschetti M, Seiger R, López HA etal (2023) A characterisation of ambiguity in bpm. In: International Conference on Conceptual Modeling. Springer, pp 277–295 Glöckner A, Herbold AK (2008) Information processing in decisions under risk: Evidence for compensatory strategies based on automatic processes. MPI collective goods preprint (2008/42) Gulden J, Burattin A, Abbad-Andaloussi A et al (2020) From analytical purposes to data visualizations: a decision process guided by a conceptual framework and eye tracking. Softw Syst Model 19:531–554 Haisjackl C, Pinggera J, Soffer P etal (2015) Identifying quality issues in bpmn models: an exploratory study. In: Int. Workshop on Business Process Modeling, Development and Support. Springer, pp 217–230 Heggset M, Krogstie J, Wesenberg H (2015) The influence of syntactic quality of enterprise process models on model comprehension. In: CAiSE Forum. pp 89–96 Hildebrandt TT (2018) Ecoknow: Effective, co-created & compliant adaptive case management for knowledge workers. In: 2018 IEEE 22nd International Enterprise Distributed Object Computing Workshop (EDOCW). IEEE, pp 9–11 Holmqvist K, Nyström M, Andersson R etal (2011) Eye Tracking: A comprehensive guide to methods and measures. OUP Oxford Jayaraman S, Phillips EK, Church D et al (2024) Privacy and utility perceptions of social robots in healthcare. Comput Hum Behav Artif Hum 2(1):100039 Keller G, Nüttgens M, Scheer AW (1992) Semantische Prozessmodellierung auf der Grundlage “ereignisgesteuerter Prozessketten (EPK)”. Institut für Wirtschaftsinformatik Saarbrücken: Veröffentlichungen des Instituts für Wirtschaftsinformatik, Inst. für Wirtschaftsinformatik Kindler E (2009) Model-based software engineering and process-aware information systems. Transactions on Petri Nets and Other Models of Concurrency II: Special Issue on Concurrency in Process-Aware Information Systems 27–45 Krogstie J (2012) Model-based development and evolution of information systems: A Quality Approach. Springer Science & Business Media Krogstie J (2016) Quality of business process models. In: Quality in business process modeling. Springer, pp 53–102 Leopold H, Smirnov S, Mendling J (2010) Refactoring of process model activity labels. In: 15th Int. Conf. on Applications of Natural Language to Information Systems. Proceedings 15. Springer, pp 268–276 Lindland OI, Sindre G, Solvberg A (1994) Understanding quality in conceptual modeling. IEEE Softw 11(2):42–49 Mendling J, Recker J, Reijers HA (2010) On the usage of labels and icons in business process modeling. Int J Inf Syst Model Des (IJISMD) 1(2):40–58 Mendling J, Reijers HA, Recker J (2010) Activity labeling in process modeling: empirical insights and recommendations. Inf Syst 35(4):467–482 Mendling J, Reijers HA, van der Aalst WM (2010) Seven process modeling guidelines (7pmg). Inf Softw Technol 52(2):127–136 Page 38 of 38Abbad-Andaloussi et al. Process Science (2025) 2:19 Mendling J, Strembeck M (2008) Influence factors of understanding business process models. In: International Conference on Business Information Systems. Springer, pp 142–153 Mendling J, Strembeck M, Recker J (2012) Factors of process model comprehension–findings from a series of experiments. Decis Support Syst 53(1):195–206 OMG (2014) Business Process Model and Notation (BPMN), Version 2.0.2. https://www.omg.org/spec/BPMN/2.0.2. Accessed 19 Dec 2024 Oppl S (2016) Articulation of work process models for organizational alignment and informed information system design. Inf Manag 53(5):591–608 Petre M (1995) Why looking isn’t always seeing: readership skills and graphical programming. Commun ACM 38(6):33–44 Petrusel R, Mendling J (2013) Eye-tracking the factors of process model comprehension tasks. In: CAiSE 2013, Proceedings 25. Springer, pp 224–239 Petrusel R, Mendling J, Reijers HA (2016) Task-specific visual cues for improving process model understanding. Inf Softw Technol 79:63–78 Petrusel R, Mendling J, Reijers HA (2017) How visual cognition influences process model comprehension. Decis Support Syst 96:1–16 Pittke F, Leopold H, Mendling J (2015) Automatic detection and resolution of lexical ambiguity in process models. IEEE Trans Softw Eng 41(6):526–544 Pustejovsky J (1998) The generative lexicon. MIT press Ralph P, bin Ali N, Baltes S etal (2021) Empirical standards for software engineering research. https://arxiv.org/abs/2010.03525. Accessed 19 Dec 2024 Reijers HA, Mendling J (2010) A study into the factors that influence the understandability of business process models. IEEE Trans Syst Man Cybern 41(3):449–462 Reinhartz-Berger I, Figl K, Haugen Ø (2017) Investigating styles in variability modeling: hierarchical vs. constrained styles. Inf Softw Technol 87:81–102 Reisig W (2012) Petri nets: an introduction, vol4. Springer Science & Business Media Schreiber C, Abbad-Andaloussi A, Weber B (2024) On the cognitive and behavioral effects of abstraction and fragmentation in modularized process models. Inf Syst 125:102424 Sennet A (2011) Ambiguity. Stanford Encyclopedia of Philosophy Sharafi Z, Soh Z, Guéhéneuc YG (2015) A systematic literature review on the usage of eye-tracking in software engineering. Inf Softw Technol 67:79–107 Slaats T, Mukkamala RR, Hildebrandt T etal (2013) Exformatics declarative case management workflows as dcr graphs. In: Business Process Management: 11th International Conference, BPM 2013, Beijing, China, August 26-30, 2013. Proceedings. Springer, pp 339–354 Sweller J (2011) Cognitive load theory. In: Psychology of learning and motivation, vol55. Elsevier, pp 37–76 Turetken O, Dikici A, Vanderfeesten I et al (2020) The influence of using collapsed sub-processes and groups on the understandability of business process models. Bus Inf Syst Eng 62:121–141 Veltman J, Jansen C (2005) The role of operator state assessment in adaptive automation. TNO Defence, Security and Safety Soesterbery Wang W, Chen T, Indulska M et al (2022) Business process and rule integration approaches–an empirical analysis of model understanding. Inf Syst 104:101901 Weber B, Fischer T, Riedl R (2021) Brain and autonomic nervous system activity measurement in software engineering: A systematic literature review. J Syst Softw 110946 Wilcoxon F (1945) Individual comparisons by ranking methods. Biom Bull 1(6):80–83 Winter M, Neumann H, Pryss R et al (2023) Defining gaze patterns for process model literacy-exploring visual routines in process models with diverse mappings. Expert Syst Appl 213:119217 Wohlin C, Runeson P, Höst M etal (2012) Experimentation in software engineering. Springer Science & Business Media Zugal S (2013) Applying Cognitive Psychology for Improving the Creation, Understanding and Maintenance of Business Process Models. PhD thesis, University of Innsbruck Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.