How do you connect?
Analysis of Social Capital Accumulation
in connectivist MOOCs
School of Interactive Arts and
Simon Fraser University
School of Informatics
The University of Edinburgh
Institute for Intelligent Systems
The University of Memphis
Schools of Education and Informatics
The University of Edinburgh
Arthur C. Graesser
Department of Psychology
The University of Memphis
School of Education
University of South Australia
Learning and Teaching Unit
University of South Australia
Connections established between learners via interactions are seen
as fundamental for connectivist pedagogy. Connections can also
be viewed as learning outcomes, i.e. learners’ social capital
accumulated through distributed learning environments. We
applied linear mixed effects modeling to investigate whether the
social capital accumulation interpreted through learners’ centrality
to course interaction networks, is influenced by the language
learners use to express and communicate in two connectivist
MOOCs. Interactions were distributed across the three social
media, namely Twitter, blog and Facebook. Results showed that
learners in a cMOOC connect easier with the individuals who use
a more informal, narrative style, but still maintain a deeper
cohesive structure to their communication.
Categories and Subject Descriptors
Education; K.3.1 [Computer Uses in Education] Distance learning
Social Processes, Automated Text Analysis, Learning
Social capital, Language, Coh-Metrix, MOOCs, Social Network
Connectivist Massive Open Online Courses (cMOOCs) scale
learner interactions by sharing, aggregating, and connecting
information through the use of a diverse set of media. This
approach allows learners to interact with each other around
personal goals and common interests, outside of the teacher-
controlled environment . However, the distributed and open
nature of cMOOCs complicates research inquiries into learning-
related processes occurring in such environments. To date, the
majority of cMOOC research has relied on self-reported
mechanisms such as course evaluations obtained through
participant surveys and identification of skills and capabilities that
effectively support learner participation [2-5].
The establishment of social ties with other learners through
interactions mediated by technology is viewed as integral to the
learning process in cMOOCs [6, 7]. The quality of the
relationships between the learners in a networked environment
can be understood through the concept of social capital .
Essentially, a large amount of social capital reflects strong and
productive relationships, based on the common interests and
shared understanding among the participants [9, 10]. In this study,
we further rely on the concept of social capital to describe the
individual learning outcomes that result from the user interactions
in cMOOCs using social media. Given that the social network
analysis focuses on the relationships between individuals, rather
than individuals and their properties , it is commonly used to
assess the social capital and estimate the opportunities and
limitations inherent to an individual actors’ position in a social
network . For example, in an analysis of Twitter-based
interactions within a cMOOC, Skrypnyk et al.  reported that
an increase in the number and density of the communication acts
resulted in an increased percentage of participants sharing the
“power and control” over the information flow with the original
course facilitators. This further means that very quickly after the
Permission to make digital or hard copies of all or part of this work for
personal or classroom use is granted without fee provided that copies are
not made or distributed for profit or commercial advantage and that copies
bear this notice and the full citation on the first page. Copyrights for
components of this work owned by others than ACM must be honored.
Abstracting with credit is permitted. To copy otherwise, or republish, to
post on servers or to redistribute to lists, requires prior specific permission
and/or a fee. Request permissions from Permissions@acm.org.
LAK '15, March 16 - 20, 2015, Poughkeepsie, NY, USA
Copyright 2015 ACM 978-1-4503-3417-4/15/03…$15.00
course started, several course participants emerged as the most
influential “factors” in knowledge sharing and brokering
information. Consequently, individual positions in a learner
network may indicate a degree of influence or faster access to
more human and technological resources in the particular course.
The present study investigated the influence of learners’ linguistic
and discourse patterns, using an automated text analysis tool, on
the accumulation of social capital. We analyzed the social
networks extracted from the learner interaction within the three
social media platforms (i.e., Twitter, Facebook, and blogs) that
were used in two cMOOCs - namely, CCK11 and CCK12, as
defined shortly. Specifically, linear mixed effects modeling
assessed the association between the accumulation of social
capital (determined through SNA) in a cMOOC and the linguistic
and discourse features  used by learners in the content created
and shared in social media.
2. THEORETICAL BACKGROUND
2.1 Social Capital
Individual positioning inside a social structure often yields
material and symbolic resources that can be organized in a
network of relationships of mutual acquaintance and recognition
. The invariable benefits of such networks are often called
social capital. In educational research, social capital has helped to
explain how the frequency and quantity of learner interactions and
relationships is related to academic performance and drop out [14,
15]. There are links between academic work, social capital theory,
and learning. For example, Gašević et al.  reported that
learners’ social capital is associated with the academic
performance, whereas Kovanović et al.  connected social
capital with the social presence in the communities of inquiry.
Social network analysis as a theory and method allows us to
translate the abstract metaphor of social capital into an observable
construct . Specifically, measuring network structural
properties, such as centrality, we are able to assess indicators
describing the social capital accumulated by each individual in the
network of learners. Besides focusing on the structural properties,
SNA also provides information about the nature of the
relationship and the strength of ties between learners. Therefore,
in this paper, the semantic meaning and value of the interactions
defines a tie between them, whereas measures of node centrality
(i.e., degree, closeness, betweenness, and eigenvalue) are used to
obtain a multi-dimensional measure of learners’ social capital.
2.2 Language Use
Researchers typically rely on structural properties, as measured
through SNA, when exploring online interactions. However, the
interactions themselves customarily take place in natural
language. From this view, language and discourse play a unique
role in computer-mediated learning environments. It is the
predominate channel used by learners to exchange thoughts and
content. Moreover, the connection between discourse and social
ties within a network is well established in numerous
anthropological, sociological and sociolinguistic studies (for a
review, ). Automated linguistics analysis methods are
particularly well suited for handling the increasing scale of
educational data. In line with this, linguistic analysis could
provide rich contextual information to the behavioral patterns
derived through SNA techniques. However, the combination of
these two analytical methods is notably scarce in the literature,
although there are exceptions . The goal of this paper is to fill
the large gap between semantic content analyses and SNA.
Specifically, we adopted a theoretically grounded, computational
linguistic analysis approach in combination with SNA to explore
student interactions within an xMOOC. Psychological
frameworks of discourse comprehension and learning have
identified the representations, structures, strategies, and processes
at multiple levels of discourse –. Five levels have
frequently been offered in these frameworks: (1) words, (2)
syntax, (3) the explicit textbase, (4) the situation model
(sometimes called the mental model), and (5) the discourse genre
and rhetorical structure (the type of discourse and its
composition). We embrace this multilevel approach to language
and discourse in the current paper.
3. RESEARCH QUESTIONS
Language and discourse is a channel of communication that is
central to the information exchange within a network of learners,
so the features of the communication channel are the main
determinants of the social and cognitive processes that evolve in
social networks . The main premise of this study is that
language and the quality of social ties established between
learners in a cMOOC are mutually dependent and correlated.
Building on this premise, we defined our research question as
RQ: How does the language used by learners in a cMOOC
influence the accumulation of the social capital?
This study examined the learner interaction occurring within
blogs, Twitter and Facebook social media from the 2011 and 2012
editions of the Connectivism and Connective Knowledge (CCK)
course. Both offerings were facilitated over a 12 week period.
Live sessions were delivered using Elluminate, while course
resources were delivered via gRSShopper. For the purpose of the
automated data collection, we used gRSShopper as the source of
links to blogs and copies of tweets. Facebook data from the
course’s open group were collected using Facebook API in order
to retrieve communication between course participants. Finally, in
order to support the analysis of content created in multiple
languages, messages posted in languages other than English were
translated using Microsoft Translation API (around 5% of the all
The total number of active learners (Ncck11=997, Ncck12=429)1 was
higher during the CCK11 course, which was also reflected in the
number of the posts created within the CCK11 (Npost11=5711,
M=2.59, SD=4.47) and CCK12 (Npost12=2951, M=3.41, SD=9.06).
However, despite a smaller cohort in the CCK12 course, the
participants demonstrated a higher average activity (Facebook:
Npost11f=1755, Npost12f=61; blogs: Npost11b=1473, Npost12b=624).
Twitter-mediated communication sustained similar high levels of
activity for both courses (Npost11t=2483, Npost12t=2266).
4.2.1 Social Network Analysis
We constructed 72 undirected weighted graphs to represent
interactions independently mediated by the three media (i.e.,
Twitter, blogs and Facebook) for each week of the two courses.
Twitter-based social networks included all the authors and
mentions as nodes of the network. The edges between two nodes
were created if an author was tagged within the tweet. For
example, if a course participant @L1 mentioned learners @L2
and @L3 in a tweet, then the course Twitter network would
contain @L1, @L2, and @L3 with the following edges: @L1-
@L2, and @L1-@L3. Social graphs from Facebook and blog
communication included authors of the posts, i.e. blog owners or
Facebook post initiators, as well as authors of comments to either
of these. Specifically, if a learner A1 created a blog or Facebook
post, and then learners B1 and C1 added comments to that post,
the corresponding network would include nodes A1, B1, and C1
with the following edges: A1-B1, and A1-C1. All the weekly
1 Number of students for courses under study, represents the number of active
students that participated in communication using three social media platforms
social graphs extracted, included authors who posted and/or
commented within the given week only.
The concept of node centrality is commonly used to assess the
importance of an individual node within the network .
Therefore, the following well-established SNA measures 
were calculated for each learner in all the network graphs: degree
centrality (i.e., the number of edges a node has in a network),
eigenvalue centrality (i.e., the measure of influence of a given
node on other nodes), closeness centrality (i.e., the distance of an
individual node in the network from all the other nodes), and
betweenness centrality (i.e., the number of shortest paths
between any two nodes that pass via a given node). For the
analyses of the social network variables we used igraph 0.7.1
, a comprehensive R software package for network analysis.
4.2.2 Linguistic Analysis
In order to conduct linguistic analysis, we parsed all the learner
generated posts across the three media in weekly chunks.
Specifically, all the posts produced by Learner 1 using Twitter as
a media, during the first week of a course, were treated as a single
unit. However, all the text produced by the same learner on
Facebook within the same week in the same course was treated as
another unit. Discourse analyses were conducted using Coh-
Metrix computational linguistic facility , . Coh-Metrix is,
arguably, the most comprehensive automated textual assessment
tool that allows for analysis of higher level features of language
and discourse , . In this study, we calculated the following
five Coh-Metrix principal components, for the each unit of
analysis: narrativity (i.e., the extent to which the text is in the
narrative genre), deep cohesion (i.e., the extent to which the ideas
in the text are cohesively connected), referential cohesion (i.e.,
the extent to which explicit words and ideas in the text are
connected with each other as the text unfolds), syntactic
simplicity (i.e., sentences with few words and simple, familiar
syntactic structures), and word concreteness (i.e., the extent to
which content words that are concrete and meaningful).
4.2.3 Statistical Analysis
All the variables (i.e., centrality measures and Coh-Metrix
principal components) were measured at the individual level, and
the data were structured in a way that learners were nested within
a course. Therefore, we adopted a mixed-effects modeling
approach, which is a recommended method for analyzing such
datasets , allowing for more stringent examination of the
effect of language on centrality by controlling for the variance
associated with individual students and course differences. Four
different linear mixed-effects models were constructed (i.e.,
centrality models), one for each of the four dependent variables:
eigenvalue, degree, closeness, and betweenness. Independent
fixed effect variables included five Coh-Metrix principal
components. Moreover, media (i.e., Twitter, Facebook, and
blogs), week, and post count were included as fixed effects.
However, given the scope of this paper, those variables are not
defined and elaborated. To address the impact of individual
variance within a model, the course and learners within a course
were treated as random effects.
The best mixed effects regression model was selected through the
several steps. Besides the model with all the fixed effects
included, null models with the random effects (student within
course, and course slope), but no fixed effects were also
constructed. A comparison of the null model with the centrality
models determined whether language predicts social dynamics
above and beyond the random effects. Intraclass Correlation
Coefficient (ICC), , Akaike Information Criterion (AIC) and a
likelihood ratio test , were used to decide on the best fitting
and most parsimonious model. An effect size (R2) was also
estimated for each model as a goodness-of-fit measure denoting
variance explained .
All the statistical analyses were conducted using R v.3.0.1
software for statistical analysis with package lme4, for fitting
linear mixed-effects models . Each of the hypotheses specify a
specific direction in the effect, therefore one-tailed tests were used
for significance testing with an alpha level of .05.
5.1 Degree Centrality
The results of the likelihood ratio test between the two models
supported the conclusion that the degree model yielded a
significantly better fit than the null model, χ2(19) = 1506.5,
p<.001. Results of the linear mixed-effect analysis (Table 1.
Centrality scores as a function of Coh-Metrix text characteristics)
revealed a significant main effect for Narrativity, F(1, 3042.7) =
4.13 p = .042, Referential Cohesion, F(1, 2806.4) = 27.32, p <
.001, Deep Cohesion, F(1, 3034.1) = 4.22, p = .040, Syntax Ease,
F(1, 3033.8) = 4.49, p = .032. Specifically, individuals with a
significantly lower degree centrality expressed themselves with a
higher degree of referential cohesion and text simplicity.
However, the learners with higher centrality scores had higher
deep cohesion and narrativity.
Table 1. Centrality scores as a function of Coh-Metrix text
Measure Degree Eigenvalue
arrativity 0.03*0.03 0.03 0.003
Word Concreteness -0.006 0.01 -0.01 0.001
Referential Cohesion -0.07*** 0.01 -0.06*** 0.001
Deep Cohesion 0.04*0.02 0.008 0.002
Syntax Simplicity -0.03*0.04 -0.009 0.003
Measure Closeness Betweenness
arrativity -0.0009 0.0004 0.02 2.82
Word Concreteness -0.02 0.0002 -0.02 1.25
Referential Cohesion 0.008 0.0002 -0.04*1.30
Deep Cohesion 0.01 0.0003 0.03 2.12
Syntax Simplicity -0.004 0.0004 -0.04*3.10
Note: * p < .05; ** p < .001. Standard error (SE). N= 3066.
5.2 Eigenvalue Centrality
Similar to the degree model, the likelihood ratio test between the
null model and the eigenvalue model revealed a significantly
better fit of the model that accounted for variation of students
within different courses (χ2(19)= 681.62, p<.001). The model
(Table 1. Centrality scores as a function of Coh-Metrix text
characteristics) showed a significant negative effect of Referential
Cohesion, F(1, 2667.4) = 13.33, p < .001. Specifically, learners
who exhibited lower scores for referential cohesion had higher
eigenvector centrality values.
5.3 Betweenness and Closeness Centrality
We initially fit the same models with respect to degree and
eigenvalue centrality to investigate how linguistic features of
computer-mediated communicative utterances predict
betweenness and closeness centrality. The models with all fixed
and random effects resulted with better overall goodness-of-fit
measures (AICc, R2, and ICC). However, further investigation of
the results for the random effects showed the perfect negative
correlation between random effects specified. This indicates that
the model overfitted the data . Therefore, we decided to
discard models with random slope and continue analysis with the
simpler models (i.e., student within a course as a random effect).
The closeness model did not reveal any significant linguistic
properties and therefore is not further elaborated. In the case of
the betweenness model, the likelihood ratio test with the null
model indicated a better fit of the model that included fixed and
random effects (χ2(19)= 390.28, p<.001). Reflecting on the
solution for the fixed effects, we were able to identify a significant
negative effect of Referential Cohesion, F(1, 3026.6) = 4.19, p =
.041 and Syntactic Ease, F(1, 3042.3) = 5.04, p = .025. The results
show that course participants who tended to use simple linguistic
constructs with higher referential cohesion had lower
6.1 Interpretation of the results
The results indicate that deep level linguistic characteristics (i.e.,
Coh-Metrix indices) influence learner interaction within a
cMOOC. This paper did not examine surface level features (e.g.,
count of posts), however it supports the claim that a systematic
and deeper analysis (beyond the surface level dialogue
characteristics) is necessary in order to obtain a more
comprehensive insight into the linguistic processes that shape
learning in network settings and influence the development of
social connections .
The results suggested that linguistic and discourse features of
written artefacts are important determinants of learning in a
cMOOC environment. Specifically, our results show that learners
whose discourse was more narrative, with deeper cohesion, more
complex linguistic structures, and low referential cohesion had
more connections, and interacted more often with their peer
learners and instructors. Likewise, learners who authored posts
with low values of referential cohesion had more ties with the
most influential, well-connected learners, as indicated with higher
values of eigenvalue centrality. Finally, higher potential for the
control of communication and brokerage of information (i.e.,
higher betweenness centrality) included learners who tend to
integrate new information (i.e., lower cohesion) within each post
and who had a discourse that was more syntactically embedded.
Thus, what is the overall effect of deep level linguistic and
discourse properties on the accumulation of learners’ social
capital in a cMOOC? Course participants who tend to use more
narrative and informal style, nevertheless still manage to maintain
a deeper cohesive structure in their communication will have
more ties. That being said, we were able to conclude that
language does define structural positions within the social
network emerging from the interaction in network learning
environment. The way the learners convey the messages and share
the information, could potentially bring them benefits in terms of
strengthening ties with peer learners and consequently increase
the social capital .
It is also indicative that individuals who created posts with higher
referential cohesion, attracted “attention” (i.e., comments and
reactions) from fewer participants. Given that referential cohesion
captures the extent to which ideas in the text are connected with
each other as the text unfolds, higher referential cohesion
indicates fewer gaps in conveying the ideas and increased text
readability and comprehension . On the other hand, referential
cohesion gaps occur when a sentence has few if any words that
overlap with previous sentences. . A possible explanation for
the relation of lower referential cohesion and possibly complex
syntactic structure with increased social capital could be
connected to the affordances of media used in the analyzed
cMOOC, i.e. Twitter and Facebook posts were noticeably shorter
than blog posts. This further implies that, in terms of overlap
between sentences and paragraphs, paragraph-to-paragraph
measures should be interpreted as post-to-post referential
cohesion. In this context, the lower referential cohesion might be
capturing a lack of overlap between an individual learner’s posts.
In this case, learners who tend to post more topically diverse
messages would naturally have less overlap and consequentially
lower referential cohesion values of their posts compared to their
more topically uniform counterparts. Therefore, it is likely that
learners who tended to provide novel information across their
posts, attracted more peers and attained more “followers”.
Likewise, low referential cohesion across the discourse might
indicate that those learners triggered many discussions about
dilemma’s and challenging topics. High referential cohesion
might indicate the redundant information across the posts with
lack of the “real” contribution to discussions and knowledge
development. Such interpretations should, however, be taken with
a degree of caution until further studies test replications and
6.2 Implications for theory and practice
Our findings have shown that learners’ ability to effectively use
language to communicate and share knowledge with peers is
essential to the building new ties and strengthening existing
connections. Moreover, being able to recognize important
information and coherently develop new ideas building on the
existing knowledge ultimately leads to the accumulation of social
capital. Finally, observed from the linguistic perspective, sharing
novel information, using concrete and coherently structured
language (i.e., written text) is perhaps the main prerequisite for
establishing new connections within the network of learners. It
seems that in highly distributed environments of cMOOCs,
learners tend to value new information, new ideas, triggering
novel potentially interesting and relevant discussions, rather than
elaborating on a single topic (or a small number of topics)
throughout a course. However, these new information have to be
comprehensive, well-structured in order to increase understanding
among learners and foster the interaction. Nevertheless, further
research is needed to assess individuals’ ability not only to
develop a social capital, but also to take advantages of the
accumulated social capital for a specific return (e.g., to facilitate
the achievement of learning outcomes).
It is questionable whether learners would be able to develop all
the necessary skills for learning in networked environment simply
by interacting with their peers. Therefore, future research needs to
investigate various instructional scaffolds and available
technological affordances that would provide guidelines for
students in developing necessary skills for learning in such
settings. Those skills, identified as “new media literacies” ,
should enable learners to use media affordances more efficiently
thus gaining more from learning in distributed learning contexts.
Eventually, changes in the way learners use the linguistic features
could provide an insight into individual’s progress in the
development of those literacies. On the other hand, as indicated by
various studies on online and distance education, personalized,
formative and timely feedback presents one of the most promising
approaches for fostering learning in online settings [35, 36].
Information gleaned from these findings suggests discourse
analytics could prove useful in creating personalized feedback for
students interacting within computer-mediated, networked
platforms in the future. For instance, a system could provide
accurate real time support for learners using an interface that
delivered suggestions via a simple pop up window or a more
sophisticated intelligent agent. Such computer-mediated support
could help course participants develop improved information
transfer and gathering skills. However, the value of such
enhancements awaits future work and empirical testing.
Nevertheless, this research might represents an initial step, by
highlighting potentially useful analytical tools, and stimulating
discussion about MOOC platforms capable of enhanced dynamic
social processing, and automated cognitive evaluation for learner
It is important to acknowledge limitations of this study. At the
time we were collecting the data for the analysis (April-August
2014), tweets posted within the both courses under study were no
longer available through Twitter API. Given that we obtained
those data using gRSShopper as a source, some of the course
interactions, such as replies, retweets, and favorites, could not be
collected. However, features such as mentions and hashtags, along
with the tweet content, were preserved. Additionally, the study
analyzed the data from a course in a specific subject domain. It is
reasonable to assume that different subject domains would be
characterized with different communication patterns. Therefore, it
would be prudent to analyze social interactions within courses
from various subject domains.
Through deep levels of text analyses, our findings show that
linguistic and discourse features have a significant impact on the
accumulation of learners’ social capital in a networked learning
setting. The findings suggest that facilitators of distributed courses
should consider a broad array of responsibilities that include and
extend simple network-formation beyond shaping and leveraging
the information flows throughout the learning network. For
example, cMOOC facilitators could introduce instructional
elements that enhance group and individual communication skills.
Finally, the study opens up for further investigation of the
relationship between social ties and language in use.
 A. A. Mcauley, B. Stewart, G. Siemens, and D. Cormier,
“The MOOC Model for Digital Practice,” p. 33, 2010.
 A. Fini, “The Technological Dimension of a Massive Open
Online Course : The Case of the CCK08 Course Tools,” Int.
Rev. Res. Open Distance Learn., vol. 10, no. 5, 2009.
 R. Kop, “The challenges to connectivist learning on open
online networks: Learning experiences during a Massive
Open Online Course,” Int. Rev. Res. Open Distance Learn.,
vol. 12, no. 3, pp. 19–38, 2011.
 R. Kop, J. Sui, and F. Mak, “A Pedagogy of Abundance or a
Pedagogy to Support Human Beings? Participant Support on
Massive Open Online Courses,” Int. Rev. Res. Open
Distance Learn., vol. 12, no. 7, pp. 1–10, 2011.
 C. Milligan, A. Littlejohn, and A. Margaryan, “Patterns of
Engagement in a Connectivist MOOC,” J. Online Learn.
Teach., vol. 9, no. 2, 2013.
 J. Anh, B. Butler, and A. Alam, “Learner Participation and
Engagement in Open Online Courses: Insights from the Peer
2 Peer University,” MERLOT JOLT, vol. 9, no. 2, 2013.
 J. Knox, “Digital culture clash: ‘massive’ education in the E-
learning and Digital Cultures MOOC,” Distance Educ., vol.
35, no. 2, pp. 164–177, 2014.
 N. Lin, “Social capital: a theory of social structure and
action,” Soc. Forces, vol. 82, no. 3, pp. 1209–11, 2004.
 A. Lockwood, “Community collaboration and social capital:
an interview with Gary G. Wehlage,” Lead. Tomorrows
Sch., vol. 2, pp. 19–25, 1996.
 J. Scott and P. Carrington, The SAGE Handbook of Social
Network Analysis. SAGE, 2011.
 R. S. Burt, “The Network Structure of Social Capital,” Res.
Organ. Behav., vol. 22, pp. 345–423, 2000.
 O. Skrypnyk, S. Joksimovic, V. Kovanovic, D. Gasevic, and
S. Dawson, “Roles of course facilitators, learners, and
technology in the flow of information of a cMOOC,” Br. J.
Educ. Technol. Spec. Issue MOOCs, p. submmited for
 A. Graesser, D. Mcnamara, and J. Kulikowich, “Coh-Metrix:
Providing Multilevel Analyses of Text Characteristics,”
Educ. Res., vol. 40, no. 5, pp. 223–234, Jun. 2011.
 C. Carceller, S. Dawson, and L. Lockyer, “Improving
academic outcomes: does participating in online discussion
forums payoff?,” Int. J. Technol. Enhanc. Learn., vol. 5, no.
2, pp. 117–132, 2013.
 L. M. Vaquero and M. Cebrian, “The rich club phenomenon
in the classroom,” Sci Rep, vol. 3, Jan. 2013.
 D. Gašević, A. Zouaq, and R. Janzen, “‘Choose Your
Classmates, Your GPA Is at Stake!’: The Association of
Cross-Class Social Ties and Academic Performance,” Am.
Behav. Sci., 2013.
 V. Kovanović, S. Joksimović, D. Gašević, and M. Hatala,
“What is the Source of Social Capital? The Association
between Social Network Position and Social Presence in
Communities of Inquiry,” in Proceedings of the Workshops
held at Educational Data Mining 2014, co-located with 7th
International Conference on Educational Data Mining
(EDM 2014), London, UK, 2014, pp. 1–8.
 A. J. Scholand, Y. R. Tausczik, and J. W. Pennebaker,
“Assessing Group Interaction with Social Language
Network Analysis,” in Advances in Social Computing, S.-K.
Chai, J. J. Salerno, and P. L. Mabry, Eds. Springer Berlin
Heidelberg, 2010, pp. 248–255.
 A. C. Graesser and D. S. McNamara, “Computational
Analyses of Multilevel Discourse Comprehension,” Top.
Cogn. Sci., vol. 3, no. 2, pp. 371–398, 2011.
 W. Kintsch, Comprehension: A Paradigm for Cognition.
Cambridge, U.K.: Cambridge University Press, 1998.
 C. E. Snow, Reading for Understanding: Toward a Research
and Development Program in Reading Comprehension.
Santa Monica, CA: Rand Corporation, 2002.
 N. Dowell, W. Cade, Y. Tausczik, J. Pennebaker, and A.
Graesser, “What Works: Creating Adaptive and Intelligent
Systems for Collaborative Learning Support,” Intell.
Tutoring Syst. Proc. ITS 2014, pp. 124–133, 2014.
 L. C. Freeman, “Centrality in social networks conceptual
clarification,” Soc. Netw., vol. 1, no. 3, pp. 215–239, 1979.
 S. Wasserman and K. Faust, Social network analysis:
Methods and applications. Cambridge, UK: Cambridge
University Press, 1994, p. 825.
 G. Csardi and T. Nepusz, “The igraph software package for
complex network research,” InterJournalournal, vol.
Complex Sy, p. 1695, 2006.
 D. S. McNamara, A. C. Graesser, P. M. McCarthy, and Z.
Cai, Automated evaluation of text and discourse with Coh-
Metrix. Cambridge, M.A.: Cambridge University Press.,
 J. C. Pinheiro and D. M. Bates, Mixed-effects models in S
and S-PLUS. Springer, 2000.
 S. W. Raudenbush and A. S. Bryk, Hierarchical Linear
Models: Applications and Data Analysis Methods. SAGE
 T. Hastie, R. Tibshirani, and J. Friedman, The Elements of
Statistical Learning, 2nd editio. New York: Springer, 2009.
 R. Xu, “Measuring explained variation in linear mixed
effects models,” Stat. Med., vol. 22, no. 22, pp. 3527–3541,
 D. Bates, M. Maechler, B. Bolker, and S. Walker, lme4:
Linear mixed-effects models using Eigen and S4. 2013.
 R. H. Baayen, Analyzing Linguistic Data: A Practical
Introduction to Statistics Using R. Cambridge University
 M. De Laat, “Networked learning,” Police Acad. Neth.
 S. Dawson and G. Siemens, “Analytics to literacies: The
development of a learning analytics framework for
multiliteracies assessment,” Int. Rev. Res. Open Distance
Learn., vol. 15, no. 4, 2014.
 R. M. Bernard, P. C. Abrami, E. Borokhovski, C. A. Wade,
R. M. Tamim, M. A. Surkes, and E. C. Bethel, “A Meta-
Analysis of Three Types of Interaction Treatments in
Distance Education,” Rev. Educ. Res., vol. 79, no. 3, pp.
1243–1289, Sep. 2009.
 E. Borokhovski, R. Tamim, R. M. Bernard, P. C. Abrami,
and A. Sokolovskaya, “Are contextual and designed
student–student interaction treatments equally effective in
distance education?,” Distance Educ., vol. 33, no. 3, pp.