Elsevier

Cognitive Psychology

Volume 86, May 2016, Pages 1-26
Cognitive Psychology

Contrasting single and multi-component working-memory systems in dual tasking

https://doi.org/10.1016/j.cogpsych.2016.01.003Get rights and content

Highlights

  • We created cognitive computer models of a dual-tasking paradigm.

  • A monolithic working memory system was contrasted against a multi-component system.

  • Model predictions were compared against both behavioral and fMRI data.

  • All evidence was in favor of the multi-component system.

Abstract

Working memory can be a major source of interference in dual tasking. However, there is no consensus on whether this interference is the result of a single working memory bottleneck, or of interactions between different working memory components that together form a complete working-memory system. We report a behavioral and an fMRI dataset in which working memory requirements are manipulated during multitasking. We show that a computational cognitive model that assumes a distributed version of working memory accounts for both behavioral and neuroimaging data better than a model that takes a more centralized approach. The model’s working memory consists of an attentional focus, declarative memory, and a subvocalized rehearsal mechanism. Thus, the data and model favor an account where working memory interference in dual tasking is the result of interactions between different resources that together form a working-memory system.

Introduction

Empirical work has shown that working memory (WM) conflicts between tasks can severely impact overall performance during multitasking (Altmann and Trafton, 2002, Borst et al., 2010, Gray et al., 2006, Jiang, 2004, Nijboer et al., 2014, Strayer et al., 2013). However, psychological theories of multitasking do not typically address how working memory is used during concurrent task performance in any detail, and consequently, how working memory conflicts can affect multitasking performance. Existing work on multitasking has either described WM as a monolithic, single-component system (Altmann and Gray, 2000, Altmann and Trafton, 2002, Best and Lebiere, 2003a, Borst et al., 2010, Fu et al., 2004, Marois and Ivanoff, 2005, Meyer and Kieras, 1997, Salvucci, 2001, Salvucci and Taatgen, 2008, Wickens, 2002, Zylberberg et al., 2010) or not at all (Aasman, 1995, Pashler, 1994, Salvucci, 2005, Schoppek, 2002). This is inconsistent with an increasing number of studies that propose differentiated WM mechanisms consisting of several subsystems, typically a focus-of-attention, an activation-based short-term memory, and modality-specific systems (Baddeley, 2000, Braver and Cohen, 2001, Cowan, 1988, Cowan, 1995, Ericsson and Kintsch, 1995, Lewis-Peacock et al., 2011, Oberauer, 2002, Unsworth and Engle, 2007, Vosskuhl et al., 2015).

In the current paper, we investigated the role of WM in concurrent multitasking. In particular, we investigated whether a single-component WM is sufficient to explain observed interference patterns in dual-tasks or whether a multi-component WM system is required. We will discuss two experiments, of which we modeled behavioral and neuroimaging results in the shape of a cognitive computer model. We show that a multi-component view of WM that includes a focus of attention, activated short-term memory, and an active rehearsal loop is able to better capture WM use during multitasking than a monolithic WM. Furthermore, the particular WM components, and consequently the interference patterns, vary depending on the particular tasks.

Classical evidence of multitasking costs comes from the Psychological Refractory Period (PRP; Telford, 1931). The PRP paradigm consists of two choice-reaction tasks, of which the stimuli are presented with a short stimulus onset asynchrony. The goal is to respond to the first stimulus (task A) before the second (task B). As the time between the onset of the first stimulus and the second stimulus becomes shorter, the reaction time (RT) for task B becomes longer. This phenomenon can be explained with the response-selection bottleneck model (RSB; Pashler, 1994). The RSB model distinguishes three phases in the component tasks of a dual-task scenario: perception, response selection, and response. The critical assumption is that perception and response can occur in parallel during a dual-task, but response selection can only be performed sequentially (Hazeltine et al., 2006, Marti et al., 2012, Pashler, 1994, Sigman and Dehaene, 2008). The RSB model has greatly influenced multitasking research, but it only addresses one particular type of task interference. It cannot, for example, explain interference effects caused by peripheral sources (Wu, Liu, Hallett, Zheng, & Chan, 2013) or memory (Hazeltine and Wifall, 2011, Strayer et al., 2013) or working memory. Working memory interference in particular can be detrimental for performance, as it does not only cause delays in task execution, but can also lead to the forgetting or misremembering of task critical information (Borst et al., 2010, Nijboer et al., 2013, Nijboer et al., 2014, Strayer et al., 2013). For example, Strayer and Johnston (2001) found that a complex phone conversation caused drivers to miss traffic signals more than twice as often.

Understanding how WM interference affects concurrent task performance requires a detailed model of the WM mechanisms themselves, as well as a good description of how these mechanisms are used within tasks. Recent WM research argues for a multi-component view of WM: for example, Unsworth and Engle (2007) show evidence for a focus of attention combined with an activated short-term memory to retrieve relevant information. Similarly Lewis-Peacock et al. (2011) distinguish the focus of attention from STM, while Vosskuhl et al. (2015) present evidence for a differentiation between WM and STM. These findings are consistent with modern theories of WM (Baddeley, 2000, Braver and Cohen, 2001, Cowan, 1988, Cowan, 1995, Ericsson and Kintsch, 1995, Oberauer, 2002). In these theories, WM subsystems include elements such as a focus-of-attention, an activation-based short-term memory, or modality-specific systems.

A multi-component WM means that task interference could occur in one or more of the mechanisms that together form WM. In addition, it means that tasks that require short-term retention or memory manipulation could use different strategies in terms of the components that are recruited to perform the task. For example, remembering the number of presentations of a certain item could be done by keeping this number in the focus of attention, or by keeping the count active in short term memory through rehearsal. However, in contrast to recent WM investigations, multitasking research has typically conceptualized WM as a single element (Altmann and Gray, 2000, Altmann and Trafton, 2002, Best and Lebiere, 2003b, Borst et al., 2010, Fu et al., 2004, Marois and Ivanoff, 2005, Meyer and Kieras, 1997, Salvucci, 2001, Salvucci and Taatgen, 2008, Wickens, 2002, Zylberberg et al., 2010). This makes it difficult to account for differences between tasks with regard to the employed WM strategy.

Evidence for different strategic approaches to multitasking was previously shown by Howes, Lewis, and Vera (2009), who used cognitively bounded rationality (CBR) analysis to show that strategies used to perform the PRP task can differ per participant. Expanding on this, Janssen and Brumby (2015) showed that participants adapt the way they perform dual tasks to external factors, such as task characteristics and incentives. These studies show that tasks can be performed using different strategies at a very elementary level of cognition, which means that a cognitive component (i.e., working memory) can be implemented in different ways across tasks to accomplish similar goals. From a concurrent multitasking perspective this means that having two tasks that use different WM strategies could lead to different performance from tasks that use the same WM strategy, as interference between tasks would occur in different mechanisms or components.

The CBR approach is one way to avoid the issue that was pointed out by Roberts and Pashler (2000): fitting theories to data does not necessarily provide empirical support for the theory. The CBR analysis allows for assessing the strengths of the support by systematically searching for the optimal rational strategy as the initial model to be tested against the data, thereby constraining the space of implementations of the theory. Our approach to avoid the modeling pitfalls identified by Roberts and Pashler (2000) was to create model predictions of both behavioral and neuroimaging data before running the experiments. This way the model’s behavior cannot be the result of overfitting, but is driven by the model’s theoretical design. In particular, we created a model beforehand that could perform the paradigm by integrating existing task models within our own modeling framework. Subsequently, we made the predictions public before performing the actual experiment.

To evaluate our WM models we integrated them into an existing cognitive architecture. This allowed us to model our paradigm in a consistent way using a framework for cognition that has been extensively tested. As stated before, current models of multitasking are unspecific in how WM is simulated. This extends to cognitive architectures in general, such as ACT-R (Adaptive Control of Thought-Rational; Anderson, 2007) and EPIC (Executive-Process Interactive Control; Meyer & Kieras, 1997). For example, EPIC assumes that working memory is not a bottleneck at all when combining multiple tasks (Meyer & Kieras, 1997). Like all cognitive resources in EPIC, different processes can access working memory in parallel (Kieras, Meyer, Mueller, & Seymour, 1999). Instead, task interference in EPIC is the result of a task-specific control strategies created by the modeler. In ACT-R the working-memory function is performed by a combination of a one-element focus of attention and an unlimited, but decaying long-term declarative memory. Both of these resources can process just a single chunk of information at a time. Salvucci and Taatgen (2008) extended ACT-R with threaded cognition, an account of how a serial architecture can account for multitasking. Instead of an explicit control strategy, threaded cognition adds a simple interleaving task-scheduling method to ACT-R: an applicable rule is picked from the task with the highest urgency, where the most urgent task is defined as the one that has least recently had a rule selected for execution. Furthermore, tasks must follow a specific etiquette: Each thread can use modules in a greedy manner, but has to release them politely. This means that a thread will use a required module as soon as it is available, and release the module as soon as its action has been completed. Given these constraints, task interference in ACT-R is the result of contention between tasks for cognitive resources, governed by threaded cognition. The memory components of EPIC and ACT-R we have discussed here show that cognitive architectures lend themselves well for monolithic WM systems, where WM content is easily accessed from a single resource. This may have led to a bias in existing multitasking models toward single-component WM systems.

In this work we investigated how working memory is used in multitasking. We focus on the complexity of, and control over, the working memory system, and how this complexity affects interference between separate tasks. In particular, we constructed an a priori computational cognitive model to examine the interactions between two working-memory tasks and one peripheral task. The model tested whether a single-component WM is able to generate accurate predictions for behavioral and neural data. We tested the behavioral predictions of this model in Experiment 1. Next, we compared neural predictions of this model to an fMRI dataset previously presented in Nijboer et al. (2014), here referred to as Experiment 2. As the single-component model could account for the fMRI data, we then compared a multi-component WM model against both the behavioral and neuroimaging data of Experiment 2. We also test the generalizability of this model by fitting it to the behavioral data of Experiment 1. The degree to which the models fit the data and generalized over datasets was used to determine whether a multi-component view of WM could better explain multitasking interference data than a single-component implementation.

In the remainder of this paper we will first explain the details of the experimental paradigm and the modeling approach. We will then proceed to discuss the a priori model, followed by the comparisons of the model against data from Experiment 1 and 2. Next, we elaborate on the changes to the model required for a better fit of neuroimaging data. We finish with a general discussion of the implications of the model for the conceptualization of WM in multitasking.

Working memory resources involved in multitasking were investigated using three different tasks: n-back, tracking, and tone-counting. During n-back, a series of letters is presented on screen. The participant is asked to indicate for each letter if it was the same as, or different from n letters ago. In terms of resources, the n-back task uses motor, visual, and WM resources (Juvina and Taatgen, 2007, Owen et al., 2005). In this paradigm we used n = 2, and in the remainder of the paper we will refer to this task as the 2-back task. In the tracking task, the goal was to keep a cursor close to a randomly moving target using a left and right button (Martin-Emerson & Wickens, 1992). Two lines that flanked the target signaled the maximum allowed error. The tracking task was predicted to use visual and motor resources, making it the only task not to use memory resources. Finally, during tone-counting a random series of low and high pitch tones were played at different intervals. The goal was to only count the high pitch tones. After a trial, participants were asked to enter the total count. The tone-counting task was expected to use aural and WM resources during the trial, and the motor resource during the response phase.

We designed a set of six conditions: the 3 single-task and all 3 possible dual-task conditions (i.e. A, B, C; AB, AC, and BC). This setup has two desirable properties. Foremost, cognitive resources can be examined through interactions between tasks, as different task combinations result in different resource conflicts. Furthermore, the possible modeling space is more constrained: models need to capture the component task behavior as well as the interactions with both other component tasks. Detailed Methods will be reported below.

We based our models1 on threaded cognition (Salvucci & Taatgen, 2008), which itself is an extension of the ACT-R cognitive architecture (Anderson, 2007). ACT-R is a general psychological theory that has been instantiated as a simulation environment in which computational cognitive models can be developed. These models allow for precise tests of the theory by forcing one to formally implement all theoretical assumptions.

ACT-R contains modules for cognitive and peripheral functions, shown in Fig. 1. Procedural memory is used to coordinate actions between these modules, using a set of if–then rules. A rule can for instance be ‘if the visual module contains a letter, then store this letter in declarative memory’. Execution of a rule takes 50 ms, and rules are executed serially. In other words, the procedural module creates a bottleneck (Anderson et al., 2005, Byrne and Anderson, 2001). While access to the remaining modules is serial as well, different modules can process requests in parallel. Threaded cognition extends the procedural system by adding the possibility to interleave the execution of production rules from different tasks: The procedural system will pick an applicable rule from the task with the highest urgency. The most urgent task is defined as the one that has least recently had a rule picked for execution. If no rule of the most urgent task matches, the next most urgent task is selected. To stop tasks from keeping modules occupied indefinitely, they are required to release a module as soon as its action has been completed. This makes it possible to have concurrently performed tasks: the production system works as a dispatcher, sending requests to various different modules and waiting for those requests to complete in order to continue. So as long as two tasks share few or no resources, the procedural bottleneck does not need to impede execution of either task, as the modules work on the tasks independently.

While ACT-R has no dedicated WM system, it has two modules that can be used as part of a WM strategy: declarative (long-term) memory and the problem state. Although the capacity of declarative memory is essentially unlimited, the chance of being able to retrieve an item decreases over time. This is implemented by giving each item an activation value, which is a numerical expression of its strength in memory. This activation value decays over time (Anderson, 2007), but increases when the associated fact is retrieved: Items that have been used more often or more recently will have a higher activation. The activation value is used during the retrieval phase, where it determines how long the retrieval will take (a higher activation value results in shorter retrievals), and whether retrieval is possible at all (a chunk can only be retrieved if its activation value exceeds a predefined retrieval threshold value).

The problem state is a buffer that can contain a single chunk of intermediate information used by a task (Borst et al., 2010). Thus, it is similar to the ‘focus of attention’ concept (McElree, 2001, Oberauer, 2009). For example, when presented with a ‘solve-for-x’ equation with two steps, an intermediate solution can be stored in the problem state. This partial solution can then be used to calculate the final answer. While information in the problem state is accessed without a time cost, replacing it takes a relatively long time (the ACT-R default value is 200 ms, see Anderson, 2007, which has functioned well to account for multitasking interference in the past, e.g., Borst et al., 2015, Borst et al., 2010). When a problem state is replaced, the old state is stored as an item in declarative memory. The problem state is therefore also the place where new declarative knowledge can be built. Although the problem state can only contain a single chunk of information, that chunk can refer to other chunks that are already in declarative memory, thereby combining several pieces of information. The main constraint is that those pieces of information are already in declarative memory. In general, there are more methods to support a working memory function besides declarative memory and the problem state: a subvocalized rehearsal loop, using hands and fingers, and, by extension, pen and paper. As a consequence, working memory use within ACT-R depends on strategy, and it is therefore a challenge for the modeler to find the appropriate strategy, as we will see in the course of this article.

ACT-R can be used to make predictions of the fMRI BOLD (Blood-Oxygen-Level Dependent) signal (Anderson, 2007), as the different ACT-R modules have been mapped to specific brain regions. These regions have been developed from literature (for an overview see Anderson, Fincham, Qin, & Stocco, 2008) and refined through a data-driven model-based approach (Borst and Anderson, 2013, Borst et al., 2015). The regions that were used for the current analysis are from the data-driven mapping from Borst et al. (2015). They selected the best fitting regions through a meta-analysis of five very different tasks, which were modeled independently from these regions. In particular interest to the current work are the regions corresponding to the problem state and declarative memory, which are part of the fronto-parietal network that is frequently associated with working memory, cognitive control, and attentional selection (Cole and Schneider, 2007, Dosenbach et al., 2007). Specifically, Borst and Anderson (2013) found that a region of the intraparietal sulcus correlated exclusively with the problem state, while the declarative module correlated exclusively with a region of the inferior frontal gyrus. This is consistent with earlier literature regarding these areas, which implicate the intraparietal sulcus in memory retrievals and updates (Bunge et al., 2002, Olesen et al., 2004, Sohn et al., 2005, Wager and Smith, 2003), and prefrontal region in episodic memory retrievals and working memory activity (Buckner and Wheeler, 2001, Cabeza et al., 2003, Dobbins and Wagner, 2005, Smith and Jonides, 1998).

Activation in the ACT-R modules can be compared to fMRI activity in the corresponding brain regions, making it possible to constrain models not only with behavioral data, but also with fMRI data (Anderson, 2007, Anderson et al., 2011). To predict the BOLD response, the activity of each ACT-R module is convolved with a hemodynamic response function that represents the BOLD response (in this paper we use the HRF proposed by SPM, which is a mix of two gamma functions; Friston, Ashburner, Kiebel, Nichols, & Penny, 2007). The result of this convolution is the activity of the module over time, expressed as a BOLD curve. Furthermore, by calculating the area under this curve we obtain the total activation for that module during a certain timespan. For a more detailed overview of this technique, see Borst and Anderson (2015).

The goal of this paper is to explore the level of complexity of working memory that is required to explain interference patterns observed during concurrent task performance. To accomplish this, we created an a priori model to generate behavioral and neuronal predictions of the paradigm. The a priori model assumed that WM interference is caused by a single bottleneck: the problem state resource (cf. Borst et al., 2013, Borst et al., 2010). These predictions were compared against experimentally obtained data. The model was used to construct a second model that uses a working memory system with multiple potential bottlenecks. In particular, the single-component working memory system of the model, which relies on the problem state, can be transformed into a multi-component system by introducing declarative memory and subvocalized rehearsal as additional mechanisms for temporary information retention. The two distinct model implementations were examined to investigate how working memory interference occurs dual task situations, and what mechanisms control interleaving of working memory processes in concurrent dual tasking.

Section snippets

Model 1

The general modeling approach that we have taken is to design models for the individual tasks, and then combine these models without the need to add anything specific for multitasking coordination (see Salvucci & Taatgen, 2008). In designing the components for Model 1, we tried to be as parsimonious as possible: working memory mechanisms were limited to the problem state and declarative memory modules. Furthermore, declarative memory is only used when the memory load of a task exceeded a single

Participants

A total of 29 participants performed Experiment 1 (19 female, Mage = 23.0, age range: 19–30). The Ethical Committee Psychology of the University of Groningen granted approval for the experiment, and written informed consent was obtained from all participants. Participants received €10 upon completion. All participants had normal or corrected-to-normal vision. We excluded one participant who did not follow the instructions correctly, and three others for performing the 2-back task at chance level,

Paradigm

Experiment 2 was previously reported in Nijboer et al. (2014), where we investigated whether neural activity found in dual tasks is simply the summation of the single-task activity, or includes an additional, dual-task specific, component. Nijboer et al. (2014) found that dual-task activation was a summation of the single-task activity, and that dual-task interference could be predicted from the overlap in active brain regions in single-task conditions, however, no formal model was provided in

General discussion

Over the course of two experiments we tested and refined a cognitive model of dual tasking that focuses on the way in which working memory is used in multitasking situations. In particular, we were interested in how interference can occur when multiple working memory tasks are performed concurrently, and what mechanisms are involved.

For Experiment 1, our model predicted that a 2-back task combined with a tone-counting task would result in the largest increase in errors, while tone counting

Acknowledgment

This research was funded by ERC-StG grant 283597 awarded to Niels Taatgen.

References (72)

  • M. Nijboer et al.

    Single-task fMRI overlap predicts concurrent multitasking interference

    NeuroImage

    (2014)
  • K. Oberauer

    Design for a working memory

    Psychology of Learning and Motivation

    (2009)
  • D.D. Salvucci

    Predicting the effects of in-car interface use on driver performance: An integrated model approach

    International Journal of Human–Computer Studies

    (2001)
  • M.H. Sohn et al.

    An information-processing model of three cortical regions: Evidence in episodic memory retrieval

    NeuroImage

    (2005)
  • N.A. Taatgen et al.

    Why do children learn to say “broke”? A model of learning the past tense without feedback

    Cognition

    (2002)
  • T. Wu et al.

    Cerebellum and integration of neural networks in dual-task processing

    NeuroImage

    (2013)
  • J. Aasman

    Modeling driver behaviour in Soar

    (1995)
  • Altmann, E. M., & Gray, W. D. (2000). An integrated model of set shifting and maintenance. In Proceedings of the third...
  • Altmann, E. M., & Trafton, J. G. (2002). Memory for goals: An activation-based model (Vol....
  • J.R. Anderson

    How can the human mind occur in the physical universe?

    (2007)
  • J.R. Anderson et al.

    Brain regions engaged by part- and whole-task performance in a video game: A model-based test of the decomposition hypothesis

    Journal of Cognitive Neuroscience

    (2011)
  • J.R. Anderson et al.

    Learning to achieve perfect timesharing: Architectural implications of Hazeltine, Teague, and Ivry (2002)

    Journal of Experimental Psychology: Human Perception and Performance

    (2005)
  • A. Baddeley

    Working memory: Theories, models, and controversies

    Annual Review of Psychology

    (2012)
  • Best, B. J., & Lebiere, C. (2003). Spatial plans, communication, and teamwork in synthetic MOUT agents. In Proceedings...
  • Best, B. J., & Lebiere, C. (2003). Teamwork, communication, and planning in ACT-R agents engaging in urban combat in...
  • J.P. Borst et al.

    Using model-based functional MRI to locate working memory updates and declarative memory retrievals in the fronto-parietal network

    Proceedings of the National Academy of Sciences

    (2013)
  • J.P. Borst et al.

    Using the ACT-R cognitive architecture in combination with fMRI data

  • Borst, J. P., Taatgen, N. A., & Van Rijn, H. (2015). What makes interruptions disruptive? A process-model account of...
  • J.P. Borst et al.

    Using data-driven model-brain mappings to constrain formal models of cognition

    PLoS ONE

    (2015)
  • J.P. Borst et al.

    The problem state: A cognitive bottleneck in multitasking

    Journal of Experimental Psychology: Learning, Memory, and Cognition

    (2010)
  • T. Braver et al.

    Working memory, cognitive control, and the prefrontal cortex: Computational and empirical studies

    Cognitive Processing

    (2001)
  • R.L. Buckner et al.

    The cognitive neuroscience of remembering

    Nature Reviews. Neuroscience

    (2001)
  • M.D. Byrne et al.

    Serial modules in parallel: The psychological refractory period and perfect time-sharing

    Psychological Review

    (2001)
  • Chavez, A., & Salvucci, D. (2003). An ACT-R model of the Wickens tracking task. Poster presented at the twenty-fifth...
  • N. Cowan

    Evolving conceptions of memory storage, selective attention, and their mutual constraints within the human information-processing system

    Psychological Bulletin

    (1988)
  • N. Cowan

    Attention and memory: An integrated framework

    (1995)
  • Cited by (20)

    • Memory-related cognitive load effects in an interrupted learning task: A model-based explanation

      2020, Trends in Neuroscience and Education
      Citation Excerpt :

      For the SA-PM model, we further generated predefined ROI-predictions [63] to inspect underlying cognitive resource demands in more detail and link our model results to predictions from Whelan [40] The learners’ motivation was not addressed explicitly in the underlying task design, which according to Whelan [40] determines brain areas related to germane cognitive load. Hence, our inspection focused only on the declarative and imaginal module, which are connected to working memory [49]. We considered them to reflect effects of task complexity as indicator of intrinsic cognitive load.

    • A step-by-step tutorial on using the cognitive architecture ACT-R in combination with fMRI data

      2017, Journal of Mathematical Psychology
      Citation Excerpt :

      Several things could be improved, as indicated above, but the basic cognitive processing is accounted for. However, in other cases we have had to abandon a first model, and propose different processing streams to account for the data—processing streams that were inspired by the neural data (Nijboer, Borst, Van Rijn, & Taatgen, 2016). In summary, the ROI analysis can be used to evaluate models because we use predefined ROIs, and do not change these ROIs on a study-by-study basis.

    • Tracking cognitive processing stages with MEG: A spatio-temporal model of associative recognition in the brain

      2016, NeuroImage
      Citation Excerpt :

      In addition to yielding new information on associative recognition, this study also highlights the limits of behavioral data and fMRI data. We have argued before that behavioral data is insufficient to constrain models of cognition (Borst et al., 2015; Nijboer et al., 2016) as it only indicates the cumulative duration of all processes (see also Anderson et al., 2016). fMRI helps by attributing different processes to different brain regions, in effect splitting a single process in to multiple processes in space.

    • On the necessity of integrating multiple levels of abstraction in a single computational framework

      2016, Current Opinion in Behavioral Sciences
      Citation Excerpt :

      Two modules are used to perceive the environment (aural, visual) and two modules can be used to make responses (manual, vocal). Furthermore, several modules process information internally: the control module (instrumental in cognitive control), declarative memory, and the problem state, a module comparable to the focus of attention in theories on working memory [19,20]. Crucially, each module adheres to a set of precisely defined properties, which are instantiated in the computational framework.

    View all citing articles on Scopus
    View full text