The Unspoken Aspect of Socially Shared Regulation in Collaborative Learning: AI-Driven Learning Analytics Unveiling ‘Silent Pauses’

Socially Shared Regulation (SSRL) contributes to collaborative learning success. Recent advancements in Artificial Intelligence (AI) and Learning Analytics (LA) have enabled examination of this phenomenon’s temporal and cyclical complexities. However, most of these studies focus on students’ verbalised interactions, not accounting for the intertwined ’silent pauses’ that can index learners’ internal cognitive and emotional processes, potentially offering insight into regulation’s core mental processes. To address this gap, we employed AI-driven LA to explore the deliberation tactics among ten triads of secondary students during a face-to-face collaborative task (2,898 events). Discourse was coded for deliberative interactions for SSRL. With the micro-annotation of ‘silent pause’ added, sequences were analysed with the Optimal Matching algorithm, Ward’s Clustering and Lag Sequential Analysis. Three distinct deliberation tactics with different patterns and characteristics involving silent pauses emerged: i) Elaborated deliberation, ii) Coordinated deliberation, and iii) Solitary deliberation. Our findings highlight the role of ‘silent pauses’ in revealing not only the pattern but also the dynamics and characteristics of each deliberative interaction. This study illustrates the potential of AI-driven LA to tap into granular data points that enrich discourse analysis, presenting theoretical, methodological, and practical contributions and implications.


INTRODUCTION
Socially Shared Regulation in Learning (SSRL) has been recognised as a pivotal framework, instrumental not only in contributing to collaborative learning (CL) success [17] but also in fostering skills for the adaptive learning process [21].At its core, regulation in learning refers to the capability of individuals and groups to purposefully respond to changing circumstances and obstacles in learning.Recent empirical findings still have shown that most learners lack the needed regulatory skills [38], a shortfall that is even more pronounced in complex collaborative contexts that characterize today's education and work landscape [3,30].The quest to understand and support SSRL has thus been a major topic of interest within the field.
SSRL is a dynamic, group-level, and cyclical regulatory process in which members engage in deliberate negotiation, interactions, and shared discourse to iteratively fine-tune and exercise metacognitive control over different aspects of their learning process (i.e., cognition, motivation, emotion, and behaviour) [17].Given the conceptualisation of SSRL as a series of processes that unfold over time, the research focus has naturally splintered across various pathways that examine its sequential and temporal mechanism.Most research in the field has primarily adopted a conversation analytic approach (i.e., coding of students' utterances) to examine the sequential manifestation of different episodes and types of learning regulation (self-, co-, socially shared-) or different dimensions of regulation [6,20].Although extensive research has been done on various dimensions of the regulation process including metacognitive, cognitive, and socio-emotional interactions [27,28], or a more recent one -deliberative interactions [12], no single study exists that has taken into account the silent pauses that are inherent in all of this verbal discourse.
Despite the absence of words, silent pauses or silence have been considered to provide important information about a speaker's internal processing [2,15].Previous studies have delineated distinct distributional patterns of pauses, each serving a specific purpose.For instance, shorter pauses often result from simple physiological necessity (respiration), some linked to semantic function, a part of discourse strategies [13,15] while longer pauses tend to reflect internal cognitive processes such as word retrieval, planning, and monitoring, as well as memory [8].One inherent complexity of SSRL is that the driving cognitive and metacognitive mechanisms of this psychological phenomenon are mental processes that often remain unspoken rather than verbalised [20].Therefore, the inclusion of these silent intervals can effectively supplement our analyses derived from utterances and may be essential to amplify our understanding beyond what is solely verbalized.
Yet, despite their potential significance, there has been a marked absence of attempts to explore the relationship of these silent pauses within SSRL research as it is a challenging undertaking for researchers to incorporate such granular data within already extensive datasets.Making sense of SSRL multi-layered processes often requires researchers to intricately code verbal utterances across multiple regulatory dimensions [6].This task's laborious requirement necessitates the need for lower code frequencies and larger analytical units, such as 30-second intervals or meaningful episode-level, effectively rendering the possibility of incorporating silent pauses.For studies that adopt a more granular, utterance-byutterance approach [27,28], the exclusion of silent pauses may be attributable to the extensive demands of line-by-line timestamping, extracting, and coding of these moments.Thus, the labour and resource constraints associated with comprehensive conversation analytics act as formidable barriers to the alignment of silent pauses and potentially other non-verbal elements.This represents a significant challenge to our methodological advancement but also a gap in our empirical understanding of SSRL processes [29].Fortunately, the evolution of learning analytics (LA) and artificial intelligence (AI)-enabled methods in recent years offers opportunities to automate the extraction and micro-annotation of silent pauses in tandem with verbal interactions [4,11].
Informed by the human-AI approach to study SSRL with the trigger concept framework and an AI-driven LA approach for examining CL process [21], this study aims to fill the abovementioned research gap by leveraging advanced LA and AI to introduce a new unexplored component of 'silent pauses' into the tracing of SSRL phenomenon.In particular, our study models groups' processes of deliberative interactions and associated silent pauses in response to cognitive and emotional trigger regulation events.In the context of (S)SRL, a trigger event can be identified as a specific incident rooted in emotional, cognitive, motivational, or behavioural aspects that hampers the advancement of a task, necessitating changes in the current learning process or enacted strategies [21].As it has previously been observed that regulation is not a frequent phenomenon in learning [25], we draw from Järvelä et al. 's [21] SSRL trigger concept framework to incorporate cognitive and emotional challenges as markers of the regulatory responses following them.Deliberation, which refers to the deliberate negotiation characteristic of interactions for regulation in CL contexts, has garnered attention in recent SSRL studies [12].We adopted this lens to examine the regulatory responses due to its alignment in granularity with those data assumptions expected in AI-enabled techniques and afford the micro-annotation of silent pauses.The main contributions of this research are answers to the following two research questions: • RQ1: What type of deliberation tactics can be identified from group interactions, including silent pauses, during the CL processes in response to trigger events?• RQ2: What were the sequential characteristics of deliberative interaction and silent pauses pertaining to different deliberation tactics?

THEORETICAL BACKGROUND 2.1 AI-driven learning analytics for examining SSRL
Collaborative learning analytics (CLA), the nexus of computersupported collaborative learning (CSCL) and LA has surfaced as an influential paradigm for understanding, diagnosing, and facilitating CL processes [40].This integration becomes particularly salient when examined in the context of SSRL in CL.The temporal aspect of (S)SRL places unique methodological demands on researchers, calling the need to consider its process, sequence, and temp orality [20,39].CLA offers innovative methods such as process mining, sequential-pattern mining, or epistemic network analysis, capable of unveiling temporally dynamic insights that are unobservable through frequency measures, to address these challenges [34].However, with its roots and fuelled in CSCL, most existing LA research focuses on investigating computer-based educational environments.The analysis relies on log data from student interactions in digital environments, followed by questionnaires and verbal documentation [10].These data types are usually subjected to descriptive and inferential statistics.This focus on digital settings overlooks the reality that the majority of CL occurs in face-to-face or blended contexts, where interactions evolve and progress in group interactions [11,24].This thus offset the increasing demand for innovative and robust analytical techniques that can keep pace with the multifaceted nature of social aspects of learning regulation.These have been reflected in recent interdisciplinary efforts to understand the complex, temporally unfolding processes of regulation through trace in other modalities such as eye-tracking, think-aloud protocols, and physiological sensors [16,22,37].Yet, a significant portion of these multimodal LA applications are oriented more towards individual self-regulated learning processes, leaving SSRL in CL relatively less explored.
Although essential for our theoretical understanding, exploratory research as such has given rise to an array of complications.Most SSRL research still relies on aggregated coded learning events to understand the learning process [14], with the expectation that these data will subsequently be triangulated with other modalities.The laborious nature of qualitative content analysis, designed to encapsulate the multifaceted regulatory processes, necessitates compromises that often result in data lacking the requisite temporal consistency and granularity for seamless alignment with other data types [29].These methodological impediments encompass challenges in data alignment, granularity, and timeliness in the application of generated inferences for instructional interventions, thereby engendering a widening rift between conventional theories and advanced LA [4].
Nevertheless, with capabilities to analyse and capture a range of data points beyond human limitations, AI-driven LA holds the promise of providing deeper insights into the temporal, dynamic, and complex nature of socially shared regulatory learning phenomena [35].

Challenges in SSRL Research
SSRL refer to the processes in which team members collectively and deliberately negotiate, adapt, and realign their group regulation processes, strategies, beliefs, and goals.The regulation phases-comprising task understanding, planning, task enactment, and reflection and adaptation-are not static but evolve in a dynamic cycle [17].This understanding of SSRL is situated within a dynamic, cyclical framework that comprises distinct phases-task understanding, planning, task enactment, reflection, and adaptation.These phases do not follow a linear pathway but are rather characterised by an iterative and cyclical adaptation of different temporalities and facets (i.e.cognitive, behavioural, motivational, and emotional) [6].
The complexity and multidimensional nature of SSRL have necessitated the need to employ advanced learning analytics, methodologies, and multimodal data to construct comprehensive models of regulation in CL [14].Yet, the majority of these studies encounter several challenges in aligning data-driven insights with the theoretical understanding of human learning [4,9], and importantly, are still only accounted for verbalised interactions.This analytical focus has overlooked the silent pauses that are inherently part of verbal interaction in SSRL settings.Research indicates that pauses have significant meanings, whether related to breathing, cognitive processing [2] such as articulatory planning, or emotional processing such as distress [36].However, the analysis of silent pauses remains largely unexplored, potentially due to the labour-intensive nature of data annotation and resource constraints.Advancements in AI-driven LA present a unique opportunity to address this gap.AI technologies have the potential to facilitate the incorporation and micro-annotation of these silent periods into the analysis of verbal interactions.This can offer a more comprehensive understanding of SSRL processes that can later be synchronised with other data modalities for a more holistic view.
In addressing the potential misalignment of data while exploring these uncharted data points of 'silent pause', this study employed a human-AI collaborative approach to study SSRL with the trigger concept framework [21] as its conceptual foundation.Järvelä, Nguyen, and Hadwin's [21] collaborative approach suggests how humans and AI can collaboratively leverage their unique strengths to explore complex phenomena and advance existing theories.In the present study, this focus is on integrating AI's computational capacity for detecting and diagnosing multimodal signals, such as silent pauses and deliberation tactics, in response to the introduced "trigger events, " alongside human expertise for contextual interpretation and validation.In doing so, this approach not only utilises the balanced distribution of human-machine resources-capacity but also ensures the close alignment between theoretical underpinnings and analytical techniques in the evolving field of education technology and learning sciences.

METHOD 3.1 Participants and learning activities
In a controlled laboratory setting, 30 secondary Finnish students (21 males, 9 females) randomly formed 10 triads and engaged in 4,452 face-to-face turns of talk while working on a shared task that lasted for 30-40 minutes.Each student is equipped with an individual laptop and collaborates using a shared Google document accessible to all participating teams.Halfway through, a simulated cognitive trigger was introduced in the form of a customer's voice message about a product allergy.Subsequently, three emotional triggers were presented at intervals of three minutes, introducing the customer's escalated negative emotional state (impatience, urgency, annoyance).Video footage of the collaboration process was captured using Insta360 Pro cameras, while high-fidelity audio was recorded via individual and central group microphones.

Data analysis
This study is informed by Järvelä, Nguyen, and Hadwin's human-AI collaboration approach [21] in combination with Ouyang's [32] AIdriven learning analytics framework, forming a three-layered structure with AI algorithms embedded across: 1) Data Pre-processing and Analysis; 2) Sequence Analysis; and 3) Pattern analysis.In the first layer, we employed an AI-enabled recording technique for the auto-segmentation and transcription of students' utterances.These were later refined by our researchers to ensure accuracy and reliability.The approach in this layer not only lessened the manual labour normally needed but also laid the groundwork for micro-annotation of silent pause and quantitative content analysis (QCA).QCA yielded codes of deliberative interactions for each talk turn.Then, in later layers, AI-driven LA methods supported by a probabilistic modelling approach were used to extract and reveal the characteristics of different deliberation tactics.These included Optimal matching (OM) for examining the similarity of groups' sequences to detect distinct clusters of deliberation tactics.Descriptive statistics provided the distribution characteristics and lag sequential analysis (LSA) were used to demonstrate the sequential characteristics of each tactic.

Quantitative Content Analysis.
The unit of analysis was a student's turn of utterances.Nevertheless, the code's analysis decision was formulated in the broader context of team dialogues to capture group-level deliberative interaction.The context window ranged, but was not limited to, from 5 to 10 turns.Each turn of the utterance was coded with one category code for deliberative interactions (see Table 1).The coding scheme was informed by recent research [12] focused on deliberative negotiation, a key adaptive mechanism posited by SSRL theories.This lens is selected for its theoretical alignment with (S)SRL framework [17] and its granularity level that is amenable to AI-enabled techniques, thereby facilitating the micromillisecond-precise timestamp annotation of 'silent pause' later and analysis.Two researchers independently coded data from two randomly selected groups, covering 20% of the total dataset.The result Cohen's Kappa score of  = 0.76 affirmed that the coding scheme provides a level of reliability that ranges from moderate to high.
As our research questions were interested in groups' regulatory responses to different trigger events, a smaller sample containing 2,125 utterances was drawn for this study.This sample covered the period starting 3 minutes before the cognitive trigger and ending three minutes after three emotional triggers, each 3 minutes apart, to ensure consistent comparison.Then, micro-annotations of 'silent pauses' at millisecond-level timestamps were added to 2,125 coded turns of discourse, thereby contributing to a dataset of 2,898 coded events.3.2.2Optimal Matching Clustering.Following initial data processing and QCA, a temporal sequence analysis was employed to explore similarities in groups' deliberative interactions, aiming to identify distinct patterns of deliberation.Each set of group interactions was partitioned into five distinct sequences corresponding to five specific phases: (1) pre-cognitive trigger, (2) after cognitive trigger, (3) after emotional trigger 1, (4) after emotional trigger 2, and (5) after emotional trigger 3. Given that groups completed their collective tasks within varying time frames, not all groups contributed data to all five sequences.Thus, the 2,898 coded events were transformed into 43 distinct three-minute segments.Then OM algorithm was deployed, employing Levenshtein distance metrics to assess the minimal cost of converting one sequence into another [1].This analysis was implemented using Python's scikit-learn library [33].The resulting distance matrix guided the application of Ward's clustering method to further categorize deliberative sequences that exhibited similar patterns across various phases.The number of clusters that provided the best fit to the data was determined by considering the Silhouette coefficient's goodness-of-fit and the dendrogram.

Lag Sequential Analysis.
To explore the sequential relationships between various deliberative interactions and corresponding 'silent pauses,' state lag sequential analysis (LSA) was performed.
After clustering groups' deliberative interactions and 'silent pauses' to specific tactics, the frequency of elements within each category was tallied.These frequencies show the preferred interactive possibilities and the deliberation strategies that participants employed within each cluster.Subsequently, transitional probabilities were calculated using overlapped sampling to determine the likelihood that certain events would follow another (or the same) activity.
Next, LSA was conducted to identify event chains that occur at frequencies greater than chance.In LSA, the term "lag" indicates the relative position of a target code in relation to a given criterion code.For instance, "lag 1" signifies that the target code immediately follows the criterion code (direct transition), while "lag 2" denotes the code positioned two steps ahead of the criterion code, and so forth (indirect transition).In this research, we explored the connections between two consecutive lags to avoid the complexity of multiple-event associations [5].Z-scores were then calculated for each transition to assess whether these transitional probabilities significantly deviated from expected values.However, given that z-scores alone do not sufficiently indicate the presence of a pattern [26], Yule's Q, a transformation of the odds ratio to a [-1 . . .+1] measuring the strength of association, was also employed.At this moment, two-event sequences of significance are ascertained.For the sequence to be classified as a significant three-event chain, it is imperative that the transition at lag 2, along with the two transitions at lag 1, display significance [31].It should also be noted that these calculations are only made when Pearson's chi-square test, validating the significance of the relationships between rows and columns in the frequency table, returned significant.The R package LagSequential [5] was applied to calculate the transition probabilities, Chi-square, z-scores, Yule's Q, and numbers of code transitions.

RESULTS & FINDINGS 4.1 What type of deliberation tactics can be
identified from group interactions, including silent pauses, during the CL processes in response to trigger events?
In response to research question one, the optimal clustering analysis identified three clusters within the 43 sequences of groups deliberative interactions and silent.Each cluster represents similar patterns of deliberative interactions and pauses that are indicative of the deliberation tactic used.Descriptive statistics of these three clusters, including sequence length, frequency, and proportion of relevant interactions for deliberation are shown in Table 2.
4.1.1Observed Tactic 1: Elaborated deliberation.As shown in Table 2 and Figure 1, the first cluster accounted for approximately 25% of the sequences but represented over one-third of the total actions.This cluster of deliberation contained a higher average sequence length than other clusters, and the biggest percentage was 'Generate options'.In this cluster, learners spent nearly 30% of their engagement in monitoring, defining problems, and discussing them with each other (i.e., 'Educate each other').It can be inferred that this group deliberation is oriented toward a higher metacognitive thinking level and potentially follows that with an elaborated and substantive argument.
The stacked bar chart in Figure 1 shows the sequences of deliberative interactions and silent pauses taken by learners in this cluster.As can be seen, silent pauses make up 18% of interactions but occur sparsely and are often brief, contrasting with longer segments like 'Define the problem' and 'Educate each other'.This indicates that these brief 'silent pauses' in this cluster' appear to function more as transitional or reflective pauses, rather than indicative of disengagement or lack of participation.It is worth noting that this cluster also highest proportion of 'Positive socioemotional interaction', 'Regulate group emo-mo', and 'Agree and implement'.This indicated that learners in this cluster employed the elaborated deliberation tactic to influence group understanding, positively influence group dynamics and form a consensus on the next action.These results reflect those of previous studies which highlight the role of reciprocal interactions in which students build on and elaborate to achieve shared understanding and deep-level metacognitive activities [23] that also support and maintain positive group dynamics [7].Among the three, this cluster maintains a consistent and moderate average sequence length of 77.8 (SD 4.01).Learners in this cluster allocate around 32% of their engagement between 'Generate options', 'Attempt ideas', 'Agree and implement' and 28.5% between 'Establish strategy', 'Monitoring', and 'Evaluating'.This deliberation tactic was used by groups to construct or attempt their solution and revise their choice through monitoring and assessment.Unlike tactic 1, which focuses on elaborated deliberation, learners in this cluster strike a balance between analysing their options and coordinating their execution attempts and agreed-upon steps.The stacked bar chart in Figure 2 further illustrates this dynamic.Silent pauses in this cluster occur more frequently and are longer in duration compared to those in the first cluster.In this context, these pauses can be inferred as either the natural silence happening during the task-execution focus period or contemplative interludes that are intrinsically linked to monitoring activities.[2,36], this pattern may also point to a more reserved group dynamic, possibly characterized by fewer back-and-forth discussions among learners.Overall, there is a clear pattern that in this cluster learners were more 'solitary oriented' and did not spend as much time on discussive deliberation.The second research question addressed the sequential associations between the type of deliberative interactions and the silent pause in each tactic cluster.To answer this question, a series of lag sequential analysis (LSA) was employed.The general anticipation of silent pauses as a common baseline in human interaction does not go unacknowledged in this study.While this pattern is also founded in our analysis, results from the LSA reveal that these assumed pauses have unique associations with different types of deliberative interactions, depending on the tactic cluster these involved.This thereby adds complexity to our understanding of group dynamics and adaptive behaviour in response to trigger events.

Sequential manifestation of different deliberation tactics .
Tactic 1: Elaborated Deliberation.'s transition probabilities matrix between different deliberation interactions and silent pauses are shown in Table 3.A chi-square test confirmed a significant relation between the rows and columns of the tallied frequencies ( 2 = 355.6,df = 121, p = 0, Monte Carlo 2-sided).It can be seen that besides those where one deliberative interaction follows by SILENT, most probable transitions are those prior to GENER-ATE ideas (e.g., EST_STRA → GENERATE, GENERATE → GEN-ERATE, NEG_SOCIO → GENERATE and those associations of higher-level thinking processes to attempts or shared agreement (e.g., EST_STRA → AGREE, EST_STRA → ATTEMPT).
Significant direct transitions (lag = 1) and indirect transitions (lag = 2) (by which z > 1.96 and Yule's Q > 0.30) are visualised in Figure 5.As can be seen, this cluster's significant direct transitions contain those from SILENT to other deliberative interactions.The only exception is REGULATE → REGULATE (z = 2.80, Q = 0.85).It might be assumed that this is due to this cluster's approach of balancing their analytical discussions with task implementation, they started to receive and monitor more negative responses from their result.This thereby invites a greater need for emotional-motivational regulation.When considering three-element sequences in LSA, only the chains AGREE → SILENT → DEFINE ( 2 = 2.19,  2 = 0.47) and DEFINE → SILENT → ATTEMPT ( 2 = 2.23,  2 = 0.49) stood out as significant.These specific sequences lend further credence to our hypothesis, indicating this cluster's higher tendency to identify problems as they execute the agreed implementation and follow that with other attempts.Based on the computation of z-values and Yule's Q, the following direct sequence (lag 1) of two elements besides those leading to SILENT was significant: GENERATE → AGREE (z = 3.67, Q = 0.52 ); DEFINE → EST_STRA(z = 2.18, Q = 0.67 ), MONITOR → EVALUATE(z = 2.01, Q = 0.47), EVALUATE → NEG_SOCIO (z = 2.03, Q = 0.78).All significant transitions are depicted graphically in Figure 4. LSA is also conducted for significant sequences of these elements.The chains MONITOR → SILENT → ATTEMPT ( 2 =2.75,  2 =0.46), AGREE → SILENT → DEFINE ( 2 = 2.00,  2 =0.43), EDUCATE → SILENT → EDUCATE, ( 2 =7.80,  2 =0.80) and SILENT → EDUCATE → SILENT ( 2 = 6.58  2 =0.56) were significant in this case.These results show that the higher-thinking processes (i.e., educating each other, defining problems, establishing strategy) in this cluster have a stronger association with their monitoring and ideation process.The LSA results corroborate the frequency statistics.Interestingly, the EDUCATE processes are punctuated by a higher frequency of SILENT pauses, suggesting a level of discomfort [36] or introspection compared to other interactions.
Tactic 2: Coordinated Deliberation.'s transitional probabilities are presented in Table 4.Likewise, a chi-square test confirmed the interdependence of rows and columns ( 2 = 494.67,df = 121, p = 0, Monte Carlo 2-sided).The transition data reveals that SILENT serves as an integral junction within this cluster's deliberative network.This central role of silent pauses suggests that they function as connecting elements, linking disparate forms of deliberation such as EST_STRA, DEFINE, MONITOR, GENERATE, AGREE, and ATTEMPT.The presence of these hubs suggests a more complex, perhaps orchestrated, form of deliberation as opposed to an elaborated or singularly focused one.
Significant direct transitions (lag = 1) and indirect transitions (lag = 2) (by which z > 1.96 and Yule's Q > 0.30) are visualised in Figure 5.As can be seen, this cluster's significant direct transitions contain those from SILENT to other deliberative interactions.The only exception is REGULATE → REGULATE (z = 2.80, Q = 0.85).It might be assumed that this is due to this cluster's approach of balancing their analytical discussions with task implementation, they started to receive and monitor more negative responses from their result.This thereby invites a greater need for emotional-motivational regulation.When considering three-element sequences in LSA, only the chains AGREE → SILENT → DEFINE ( 2 =2.19,  2 =0.47) and DEFINE → SILENT → ATTEMPT ( 2 =2.23,  2 =0.49) stood out as significant.These specific sequences lend further credence to our hypothesis, indicating this cluster's higher tendency to identify problems as they execute the agreed implementation and follow that with other attempts.5. Again, a chi-square test substantiated the interdependence between rows and columns ( 2 = 781.96,df = 121, p = 0, Monte Carlo 2-sided).Notably, the primary patterns of transitions involve sequences to and from SILENT, but with increased probabilities and strength of association compared to Tactic 2, underscoring the more 'solitary' nature of the deliberative tactics employed in this cluster.In this study, we adopted AI-driven LA with a human-AI collaboration approach to examine deliberative interactions with silent pauses for SSRL.Most recent research has made efforts to address methodological challenges in examining SSRL, including triangulating multimodal data and the application of LA techniques [19,25].Although serving well in illuminating the temporal and cyclical processes of SSRL, most of these focus on the rich layer of verbalised interaction whilst ignoring the 'silent pause' that is not only inherent to those but embodies numerous internal processes of learners.These approaches may therefore fall short of generating insights bridging our holistic understanding of this multi-level, multi-facet, dynamic and largely unobservable psychological process at the core of regulation.Our study is one of the first attempts to incorporate 'silent pause' into an AI-enabled LA-driven examination of groups' deliberative interactions in the context of face-to-face CL at a granular level.
Previous studies have underscored learners' tactical and strategic choices in responding to situational needs and challenges as important markers of successful learning [7,17].Therefore, the goal of our study was to address two major research questions, with the aim of understanding whether we can identify meaningful deliberation tactics in response to trigger events across different groups.To that end, three distinct deliberation tactics became apparent among groups in response to regulation trigger events.The tactics detected were defined as Elaborated deliberation, Coordinated deliberation, and Solitary deliberation.The Elaborated deliberation tactic is characterized by extended group conversations centred on higher cognitive processes such as analysis and reasoning.These discussions exhibit limited 'silent pauses, ' indicating active engagement and attuned contribution.In contrast, the Coordinated deliberation is marked by a distinct nature of action-oriented discourse, where 'silent pauses' often denote stages of task execution or brief moments of reflective monitoring.The Solitary deliberation, on the other hand, features a more fragmented conversational flow, with frequent and prolonged 'silent pauses', hinting at uneven group engagement and individual contemplation.These results are consistent with prior research, which has identified varying levels of participation [18], depth and sharedness of conversation [23] across groups when examining SSRL.Moreover, these aspects have been identified as predictors for the emergence and quality of SSRL.However, the empirical, data-driven, process-oriented modelling of the micro-processes that underlie these patterns of previous work was limited.
Furthermore, in an advancement over prior research, our study leverage AI-driven LA techniques to incorporate 'silent pauses' into the analysis, thereby shedding new light on the deliberation process that group undertake in CL.Although previous approaches have effectively provided insights into the overall trajectory or sequential process of externalised verbal interactions for SSRL, this study's inclusion of 'silent pauses' which as previous studies have established -carry indicators of internal cognitive and emotional processing [15,36] brings two additional layers of understanding to the table.First, it explicates the group dynamics through the discovery of three distinct deliberation tactics as discussed above.Second, it adds a temporal dimension to each verbal interaction, enabling an in-depth yet modellable analytical interpretation into its nature.For example, the occurrence of frequent 'silent pauses' in a window of the 'Educate each other' episode may point to cognitive overload affecting the articulation of reasoned arguments [2].Alternatively, it could signal a fragmented or disengaged interaction, perhaps revealing minimal interest in group members expressing ideas to align shared grounds.Our study not only provides the first understanding of 'silent pauses' in revealing complex and dynamic constructs of a group's deliberative interactions for SSRL but also emphasises the importance of using it to further our holistic understanding of this complex phenomenon.In doing so, this research lays the groundwork for bridging the insight from verbalised discourse and its intertwined non-verbalised internal mental process that is at the core of regulation in CL contexts.
The proposed method and results of this study offer important implications for research and practice for different educational stakeholders.First, our methods address the current methodological gap in the field.Despite its significant connotation of internal cognitive and socio-emotional processes; the inclusion of 'silent pauses' in a micro process-oriented analysis has been limited.Adopting a human-AI collaboration approach, this study demonstrated how AI-driven LA can be used to capture a broader spectrum of previously unobtainable data points.Recent discourse in the field has raised concerns about the risks of theoretical gaps in recent research exploring advanced LA techniques [9,10].Our methods not only yield robust analyses aligned with SSRL theory but also advance the operationalization of additional analytical algorithms and future research incorporating multimodal data.Second, our methodology can enable the real-time tracing and identification of group deliberation tactics.This can facilitate early predictions of group dynamics, patterns of regulation response which can, in turn, be used to inform (i.e., via LA Dashboards) and motivate a student's change toward tactics that have proven effective in a similar context.The same could also be provided to support teacher-directed feedback and interventions.These implications can extend significantly to the design of intelligent tutoring systems and other educational technologies aimed at fostering adaptive learning environments.
Our study is not without limitations.First, the sample size is limited which renders our finding's generalisability.Second, although our approach may take us a step closer to revealing the internal process of regulation, it does so within the constraints of a limited body of existing research on 'silent pauses' in the SSRL context.Thus, these results serve more as a preliminary exploration rather than definitive conclusions and should be interpreted cautiously until further substantiated by larger-scale studies and a more comprehensive theoretical integration of 'silent pauses.'

Figure 1 :
Figure 1: Frequency of different deliberative interactions and silent pauses & state sequences of tactic 1; Note: each horizontally stacked bar represents a sequence with colours representing the type of interaction, and lengths correspond to durations

4. 1 . 2
Observed Tactic 2: Coordinated deliberation.As shown in Table2and Figure2, the second cluster accounted for approximately 28% of the sequences and represented just over one-third of the total actions.

Figure 2 :
Figure 2: Frequency of different deliberative interactions and silent pauses & state sequences of tactic 2

4. 1 . 3
Observed Tactic 3: Solitary deliberation.Compared to clusters one and two, cluster three was relatively much shorter with an average length of 48.90.This tactic was also commonly used by learners (46.51% of total sequences).The most prominent characteristic of this cluster was that in the shorter time period (average duration Cluster 3 = 165.34s;Cluster 1 = 194.87s,Cluster 2 = 170.31s)learners spent more, exceeding one-third of their engagement (35.99%) to moments of silence.Furthermore, as illustrated in the stacked bar chart presented in Figure3, these silent pauses exhibit a heightened occurrence rate and a longer temporal duration compared to the other clusters.While research in speech and language have posited that longer silent pause could indicate a focus on deep individual reflection or internal cognitive processing

Figure 3 :4. 2
Figure 3: Frequency of different deliberative interactions and silent pauses & state sequences of tactic 3

Figure 4 :
Figure 4: Tactic 1 state transition diagram of significant transitions (z > 1.96, Q > 0.30) occurring in the deliberative interaction data.Edges are labelled with their z-score.Lag 1 is continuous line; Lag 2 is dash line.

Figure 5 :
Figure 5: Tactic 2 state transition diagram of significant transitions (z > 1.96, Q > 0.30) occurring in the deliberative interaction data

Figure 6 :
Figure 6: Tactic 3 state transition diagram of significant transitions (z > 1.96, Q > 0.30) occurring in the deliberative interaction data

Table 1 :
Coding scheme for Deliberative Interactions S1: Shall we put kale in there when?S2: It sounds a bit strange [. . .] S1: Now it's good.Wise one about 500 [Complement group's strategy] Agree and implement AGREE Confirm shared agreement on the options, ideas, and opinions and carry it out.S1: Yeah, I'll change them to one hundred and twentyfive.[Inform current process from previous agreement] S2: If it's the same for you, then we'll trust it.S1: Well, if only we scored something.[All group members non-verbal show a lack of motivation]

Table 2 :
Descriptive statistics for three clusters detected with the OM algorithm