This document outlines research on modeling socio-cognitive learning processes in collaborative environments. It discusses:
- The interplay between primary memory (attention control) and secondary memory (long-term memory) during reflection on resources.
- Three studies examining how semantic stabilization evolves through this interplay and affects individual learning. Stabilization was found to support learning by priming memory searches.
- A computational model called CMR that formalizes these memory dynamics and allows modeling how people's reflections lead to intersubjective understandings and semantic stabilization in collaborative tagging systems.
2016-05-27 Venia Legendi (CEITER): Paul Seitlinger
1. The role of primary and secondary memory
in organism-environment dynamics
...
computa(onal modeling of accruing data in collabora(ve
learning scenarios
paul seitlinger
27. 05. 2016, tallinn
1
2. Outline of my research
• Knowledge building in open/self-directed learning seDngs
• Challenges from a psychological perspec(ve
– Organism-Environment dynamics
• Interplay of primary memory (scope and control of aIen(on) and
secondary (long-term) memory during reflec(on
– Secondary memory (SM): Integra(ng episodic memory (evolving in collabora(ve learning
seDng) into seman(c (pre-exis(ng) memory [1]
– Primary memory (PM): controlled use of contextual cues (environmental, internal) to
search secondary memory (interpreta(on, reflec(on) [1]
– Socio-cogni(ve processes/co-crea(on of environments
• How stabiliza(on (paIerned prac(ces, grounding) evolves and affects PM-
SM interplay
[1] Usworth, N. & Engle, R. (2007). The nature of individual differences in working memory capacity: Ac(ve maintenance in
primary memory and controlled search of from secondary memory. Psychological Review, 114, 104-132.
2
3. Outline of my research
• Observing socio-cogni(ve learning in Web environments
– Collabora(ve learning so`ware (school, university)
– Social informa(on systems (bookmarking, tagging, informa(on search)
• Making use of accruing datasets to validate models of socio-cogni(ve learning
• 3 studies by the example of social tagging
– Study 1: Field study (university course) about effects of seman(c
stabiliza(on on individual learning [2]
– Studies 2 and 3: Measurement and computa(onal modeling to shed light
on variables giving rise to stabiliza(on [3,4]
[2] Ley, T. & Seitlinger, P. (2015). Dynamics of human categoriza(on in a collabora(ve tagging system:
how social processes of seman(c stabiliza(on shape individual ssensemaking. Computers in
Human Behavior, 51, 140-151.
[3] Seitlinger, P., Ley, T. & Albert, D. (2015). Verba(m and seman(c imita(on inindexing resources on the
Web: a fuzzy-trace account of social tagging. Applied Cogni?ve Psychology, 29, 32-48.
[4] Seitlinger, P. & Ley, T. (2016). Reconceptualizing imita(on in social tagging: a reflec(ve search model
of human web interac(on. In P. Parigi & S. Staab (Eds.), Proceedings of the 8th Interna?onal ACM
conference on Web Science Conference (in press). New York: ACM press. 3
5. Study 1: Effects of seman(c stabiliza(on on individual learning
• N=24 students of a university course on cogni(ve models in TEL
– Social bookmarking system (SOBOLEO) to collect and tag Web resources
– Training phase to become familiar with purpose of tagging
• Manipula(ng stabiliza(on (λ) of tag vocabulary (high vs. low stabiliza(on)
– Low λ (n=12): ‘Old’ and interfering tags of training phase remain in the system
– High λ (n=12): Environmental switch
• Elici(ng individual learning: Performing an aIribute lis(ng task
– Lis(ng aIributes to general, medium, and specific tags (level of specificity)
• Basic-level shiN (e.g. [7])
• Hypothesis: Individuals of the high λ group gain more knowledge about medium
and specific tags than individuals of the low λ group.
[7] Close, J. & Pothos, E. (2012). “Object categoriza(on: Reversals and explana(ons of the basic-level advantage” (Rogers
& PaIerson, 2007): A simplicity account . Quarterly Journal of Experimental Psychology, 65, 1615-1632.
5
6. Study 1: Effects of seman(c stabiliza(on on individual learning
0 50 100 150
010305070
Consecutive tag assignments
Numberuniquetags
λ high
λ low
N = H * (1 – e-λt)
λ high = .009
λ low = .006
Stabiliza(on on group level: Higher
stabiliza(on in high than in low λ group
12345
Specificity
Numberlistedattributes
General Medium Specific
λ high
λ low
Individual learning: More knowledge about
medium and specific tags (basic-level shi`) in
high than low λ group
[2] Ley, T. & Seitlinger, P. (2015). Dynamics of human categoriza(on in a collabora(ve tagging system: how social
processes of seman(c stabiliza(on shape individual ssensemaking. Computers in Human Behavior, 51, 140-151.
F2,21 = 5.06, p < .05
6
7. Study 1: Effects of seman(c stabiliza(on on individual learning
• Stabiliza(on during collabora(on supports learning
– Tag-based (seman(c) priming inves(gated
by [6]
– Goals of studies 2 and 3: Revealing
interplay of remaining variables
• Study 2: Measuring i) contribu(ons of
PM-SM interplay and ii) impact of
intersubjec(vity on imita(on
• Study 3: Computa(onal model of
mechanisms underlying these variables
[6] Fu, W.-T., Kannampallil, T., Kang, R. & He, J. (2010). Seman(c imita(on in social tagging. ACM Transac?ons on
Computer-Human Interac?ons, 17, 12:1-12:37.
Tag
(contextual cue)
Search of memory
(PM-SM interplay)
Semantic priming
Reflection
(PM-SM interplay)
Intersubjectivity
Imitation
Semantic
stabilization
7
8. Study 2: Measuring the impact of intersubjec(vity on imita(on
• Web-based experiment
• 48 students conduc(ng an
informa(on search
• Incidental learning: Browsing
pictures (taken by famous
photographers; e.g., Henri Car(er-
Bresson) interpreted and
annotated by tag clouds
• Tagging phase: Re-exposed to
pictures to reflect on it and derive
own interpreta(ons and tag
assignments
• Frequency distribu(ons for the act of imita(ng (I) vs. not imita(ng (N) previously seen tags
• Analysis in terms of theore(cal constructs (e.g., PM, SM, intersubjec(vity) through
Mul(nomial Processing Tree (MPT) derived from Fuzzy-Trace Theory (e.g., [8])
[8] Brainerd, C. & Reyna, V. (2010). Recollec(ve and non-recollec(ve recall. Journal of Memory and Language, 63, 425-445. 8
9. Study 2: Measuring the impact of intersubjec(vity on imita(on
Automatic unloading
from PM
Reflective search
of memory
PM-SM interplay
Similar reflection
(Intersubjectivity)
1-S
Same tag choice
1-C
1
1-C
2
Imitation, VI
Web
resource
No Imitation, N
Imitation, I
D
1-D S
C
1
C
2 Imitation, I
No Imitation, N
Same tag choice
Different
reflection
[3] Seitlinger, P., Ley, T. & Albert, D. (2015). Verba(m and seman(c imita(on inindexing resources on the Web: a fuzzy-
trace account of social tagging. Applied Cogni?ve Psychology, 29, 32-48.
• Performing maximum likelihood es(ma(on to test model fit and quan(fy contribu(on of
cogni(ve processes
9
10. Automatic unloading
from PM
Reflective Search
of memory
PM-SM interplay
Similar reflection
(Intersubjectivity)
1-S=0.81
Same tag choice
1-C
1
=0.37
1-C
2
=0.97
Imitation, I
Web
resource
No Imitation, N
Imitation, ID=0.15
1-D=0.85 S=0.19
C
1
=0.63
Imitation, I
No Imitation, N
Same tag choice
C
2
=0.03Different
reflection
Model fit, G2(4)=0.78, n.s.
P(I)=0.10
P(I)=0.02
Probability
P(I)=0.15
Study 2: Measuring the impact of intersubjec(vity on imita(on
• PM-SM interplay crucial to model students’ interpreta(ons and annota(ons
• Intersubjec(vity (state of reflec(ve agreement) as a driving force behind imita(on
and thus, stabiliza(on
[3] Seitlinger, P., Ley, T. & Albert, D. (2015). Verba(m and seman(c imita(on inindexing resources on the Web: a fuzzy-
trace account of social tagging. Applied Cogni?ve Psychology, 29, 32-48.
10
12. Context Maintenance and Retrieval Model (CMR [9])
PM-SM interplay when reflec(ng on environmental objects
Article about
learning and
memory
“Brain” “Synapse”
“Kandel”
Item layer F
Context layer C
Context evolution (internal spotlight)
Episodic learning (integration of item-context associations into MFC
and MCF
MFC
MFC MFC
MCF
MCF MCF
Stream of thoughts triggered by environmental item
* PM: Turning environmental cues into context
** Using context to search SM
[9] Polyn, S., Norman, K. & Kahana, M. (2009). A context maintenance and retrieval model of organiza(onal processes in
free recall. Psychological Review, 116, 129-156.
*
**
12
13. Study 3: Applying CMR to model students’ reflec(ons on Web
resources as a PM-SM interplay
• CMR: A valid model of PM-SM dynamics
– Tested by a series of laboratory experiments on episodic learning (e.g., [9,10])
• RQs: Does PM-SM dynamics formalized by CMR allow for modeling
– peoples’ reflec(ons on Web resources?
– the effect of intersubjec(vity on seman(c stabiliza(on?
• RQs inves(gated in a large-scale social tagging system (Delicious)
– Dataset [11]: 1,685 tags for 49,691 Bookmarks of 2,003 Wikipedia ar(cles
from 1,968Users
• Tes(ng a CMR-specific hypothesis about stabiliza(on (consensual tag use)
• Simula(ng empirical paIerns by means of a CMR-based mul(-agent
simula(on (MAS)
[9] Polyn, S., Norman, K. & Kahana, M. (2009). A context maintenance and retrieval model of organiza(onal processes in free
recall. Psychological Review, 116, 129-156.
[10] Healey, M. & Kahana, M. (2016). A four component model of age-related memory change. Psychological Review, 123, 23-69.
[11] Zubiaga, A. (2009). Enhancing naviga(on on wikipedia with social tags. In Wikimania 2009. Wikimedia Founda(on, 2009.
13
14. Hypothesis: Decreasing intersubjec(vity during reflec(ons
F
C
F
C
F
C
F
C
Evolving spotlight
tag1
tag2
tag3
tag4
Tag assignment TAS
• TAS as a manifesta(on of
resource reflec(on (Study 2)
• Each search itera(on yields a
single tag (posi(on t) within
TAS
• Dri`ing spotlight hypothesis
• The longer we reflect, the more individualis(c the spotlight (internal context
state) should be
• Intersubjec(vity should decrease along consecu(ve search itera(ons t
(TAS posi(ons)
à Less imita(on and thus, seman(c stabiliza(on (consensual tag use)
at later TAS posi(ons
14
15. 0.40.50.60.70.80.91.0
Probabilitynewtag
Consecutive TAS
1 2 3 4 5 6 7 8 9 10
With each new TAS, the probability of a new
tag declines ~ Stabiliza(on
Web
resource
TAS
1
= {Kandel, brain, synapse, learning}
TAS
2
TAS
3
TAS
10
…
Micro dynamics
Macro
dynamics
TAS…Tag assignment
Indica(on of intersubjec(vity, implicit agreement on conceptualizing an object (e.g.,[12] )
Criterion: Seman(c stabiliza(on in a social tagging system
[12] S. Sen, S., Lam, S., Rashid, A., Cosley, D., Frankowski, D., Osterhouse, J., Harper, F. & Riedl, J. (2006). Tagging, communi(es,
vocabulary, evolu(on. In Proc. 20th anniversary conference on Computer Supported Coopera(ve Work (pp. 181- 190). ACM press. 15
17. …
…
MFC
MCF
1) Category combination of present Wikipedia article fi
3) Context evolution
ci
= ci-1
+ β*cIN
2) Context retrieval
cIN
= MFC
fi
4) Activation of
item layer
fIN
= MCF
ci
Semantic
Pre-exist.
Episodic
Evolving
(1 − !)!!"#
!"
+ !!!"#
!!
Study 3: Applying CMR to model students’ reflec(ons on Web
resources as a PM-SM interplay
• MAS
– Each agent behaves according to CMR model
1) Training phase based on a real user history (sequence of bookmarked ar(cles)
Developing individual stream of consciousness (episodic learning and spotlight evolu(on)
2) Tagging phase: All agents assign 4 tags to each of 10 further ar(cles
Semantic utility
based on reflection
u(w) = p(w|fIN
)
u’(w) = u(w)[1+s(w)]Φ
O E
Environmental salience
based on previous TAS
s(w) = p(w|fi
)
5)
• Gene(c algorithm
exploring parameter space
• 500 simula(on runs with
best-fiDng parameter set
[4] Seitlinger, P. & Ley, T. (2016). Reconceptualizing imita(on in social tagging: a reflec(ve search model of human web interac(on.
In P. Parigi & S. Staab (Eds.), Proceedings of the 8th Interna?onal ACM conference on Web Science Conference (in press). New York:
ACM press. 17
18. Study 3: Applying CMR to model students’ reflec(ons on Web
resources as a PM-SM interplay
0.40.60.81.0
Probabilitynewtagpnew(r,t)
Consecutive TAS r
Data
CMR
1 2 3 4 5 6 7 8 9 10
t = 1
0.40.60.81.0
Probabilitynewtagpnew(r,t)
Consecutive TAS r
Data
CMR
1 2 3 4 5 6 7 8 9 10
t = 3
0.40.60.81.0
Probabilitynewtagpnew(r,t)
Consecutive TAS r
Data
CMR
1 2 3 4 5 6 7 8 9 10
t = 4
• Model fit: χ2(29) = 13.74, χ2
crit=42.56
– CMR-based modeling of reflec(ng on resources
explains paIerns qualita(vely and quan(ta(vely
• Dri`ing spotlight hypothesis HDS
– Slope λ of pnew(r,t) along consecu(ve r decreases
with increasing t
Data CMR
pnew λ pnew λ
t = 1 .580 .093 .584 .089
t = 2 .639 .077 .633 .078
t = 3 .669 .069 .665 .069
t = 4 .708 .060 .700 .064
0.40.60.81.0
Probabilitynewtagpnew(r,t)
Consecutive TAS r
Data
CMR
1 2 3 4 5 6 7 8 9 10
t = 2
[4] Seitlinger, P. & Ley, T. (2016). Reconceptualizing imita(on in social tagging: a reflec(ve search model of human web interac(on.
In P. Parigi & S. Staab (Eds.), Proceedings of the 8th Interna?onal ACM conference on Web Science Conference (in press). New York:
ACM press. 18
19. Tag
(contextual cue)
Search of memory
(PM-SM interplay)
Semantic priming
Reflection
(PM-SM interplay)
Intersubjectivity
Imitation
Semantic
stabilization
Conclusion
• A valid model of peoples’ reflec(ons on Web resources
– PM-SM interplay (spotlight-driven search of memory)
• Precise predic(ons and modeling of stabiliza(on
– By implemen(ng result of study 2: Imita(on as an epiphenomenon of
intersubjec(vity (state of reflec(ve agreement)
• Studies 1-3 as a triangula(on of
• Field experiment: Iden(fying mutual
influences between observable variables on
group and individual
• Mul(nomial modeling of Web-based
experiments: Quan(fying contribu(ons of
latent variables to observable behavior
• Mul(-Agent Simula(on: Tes(ng assump(ons
on dynamics between mul(ple latent and
observable variables
19
20. Methodological implica(ons
• Collabora(ve learning scenario well captured by nonlinear organism-
environment dynamics
– No simple cause-effect rela(onships [13]
– Non-linear processes and mutual influences between variables
• Methodological approach
– Going beyond correla(onal analysis
– Computa(onal modeling
• Model-based simula(ons/predic(ons of system development
• Model-based representa(on and computa(on/analysis of
contextual informa(on about a student (temporal, seman(c,
social)
[13] Larsen-Freeman, D. Cameron, L. (2008). Research methodology on language development from a complex systems perspec(ve.
The Modern Language Journal, 08, 200-213.
20
21. Going beyond correla(onal analysis
• Advantage of computa(onal modeling
– Close to phenomena to be observed
• Distribu(on of informa(on through non-linear and itera(ve processes
– Intertwining theory and sta(s(cs
• Parameters directly represen(ng theore(cal constructs
– Independence of domain and data
• Fundamental mechanisms of
– learning (Hebbian learning of seman(c and episodic associa(ons)
– execu(ve func(ons (scope and control of aIen(on ~ Spotlight and spotlight-
driven search)
• should account for different behavioral data
– Self-directed naviga(on (forma(on of informa(on goals, meta-cogni(ve
processes/control)
» spotlight -> informa(on goal
» PM-SM interplay to account for meta-cogni(ve processes
– Crea(ve group cogni(on (trade-off between stabiliza(on and divergent
thinking)
» Interplay of aIen(on control and scope of aIen(on when re-combining
pre-exis(ng associa(ons
21
22. Contribu(ons to research infrastructure
• Summerschool on theory-driven analyses of human-web interac(ons
– Prof. Wai-Tat Fu (Partner in two currently running FWF projects)
• Department of Computer Science, University of Illinois at Urbana-Champaign
– Topic1: Computa(onal modeling of user behavior in crea(ve and self-directed learning
environments
– Topic 2: Design of crea(vely s(mula(ng recommenda(on mechanisms: „Escaping the
echo chamber“
• Summerschool on web-based experiments on „access to knowledge“
– Prof. Harry Bahrick (Partner in a current EU project proposal)
• Department of Psychology, Ohio Wesleyan University
– Topic: Applying MPTs to analyze learning in Web-based experiments
• Availability vs. Accessibility of knowledge
22