Title: Building Students Metacognitive Skills through Interactions with Computer-based Teachable Agents
1Building Students Metacognitive Skills through
Interactions with Computer-based Teachable Agents
- Gautam Biswas
- gautam.biswas_at_vanderbilt.edu
- Dept of EECS ISIS
- Vanderbilt University
- Collaborators Dan Schwartz, Kefyn Catley
- Postdoc, Students (at Vanderbilt) Rod Roscoe,
John Wagster, Hogyeong Jeong, Nancy Morabito, Jim
Segedy, Garrett Linn - Supported by Dept. of Education IES, and NSF
REESE Awards
2Goals of our work
- Learn Science through principles that apply
across domains - Processes, Entities, Relations, Interdependence,
and Balance - Preparation for Future Learning
- Students should become independent learners, even
when they move away from the computer environment - Learning for oneself ability to assess ones
learning progress - Learning is never a one step process
- Cognitive tasks and Metacognitive strategies
3Outline of Talk
- Our Approach to Learning by Teaching
- Bettys Brain, a Teachable Agent
- Learning Science by creating Causal Concept Maps
- Assessment through self-other monitoring
- Adaptive Tutoring
- Providing Metacognitive support in support of
Preparation for future learning - Experimental Studies
- Current/Future Work
4Bettys Brain
Additional resources
- Betty can explain her answers
5Teachable Agents
- Students teach computer agent using visual
representations - Agents performance based on what she is taught
- Students re-teach agent so that they may do
better, (and in that process they learn)
- Agents only learn what they are taught explicitly
by student - No machine learning algorithms drive our agent
- Learning through social interactions
- Shared representations
- Shared responsibility
6Learning Science
- By creating visual concept map structures
- Entities
- e.g., fish, macroinvertebrates, dissolved oxygen
- Relations
- causal fish consume macroinvertebrates
- increase decrease effects
- Causal Reasoning
- Cause-effect relations extended to chain of
events - Fish ? waste ? bacteria ? nutrients ? algae
- Interdependence
- Multiple dependencies everything depends on each
other
7Metacognition to aid Learning
- Metacognition describes two component processes
- Ability to monitor ones cognitive activities
- Ability to take appropriate regulatory steps when
problem is detected - Implemented as Self-regulated learning strategies
- Involves multiple aspects when learning
- Setting goals
- Planning
- Seeking help
- Monitoring ones own learning
- .
8Monitoring when Problem Solving
- Self monitoring (cf. to self explanation)
requires two coordinated processes - Ability to generate solution steps
- Analyze and correct for discrepancies
- Our approach Self-other monitoring while
teaching ( learning for oneself) - Provide support to help students organize their
own learning - Betty demonstrates self-regulated learning
behaviors by example - Mentor provides additional support and hints
9Example regulation strategies
Regulation Goal Pattern Description Bettys response
Monitoring by asking Queries Successive quiz requests but no queries asked of Betty in between quizzes Im still unsure of this material and I would like to do well. Mr. Davis said take the quiz only if you think you will do well. (Betty refuses to take quiz)
Monitoring through Explanations Multiple requests for Betty to give an answer but no request for explanation Lets see, you have asked me a lot of questions, but you have not asked for my explanations lately. Please make me explain my answers so you will know if I really understand.
Tracking Progress The most recent quiz score is significantly worse than the previous score I would really like to do better. Please check the resources, teach me, and make sure I understand by asking me questions that are on the quiz. My explanation will help you find out why I am making mistakes in my answers. Also, be sure to check out the new tips from Mr. Davis.
10Example regulation strategies
Regulation Goal Pattern Description Mentors response
Monitoring through Explanations Multiple requests for Betty to give an answer but no request for explanation Without asking Betty to explain her answers, you may not know whether she really understands the chain of events that you have been trying to teach her. Click on the Explain button to see if she explains her answer correctly.
Tracking Progress The most recent quiz score is significantly worse than the previous score Betty did well on the last quiz. What happened this time? Maybe you should try re-reading some of the resources and asking Betty more questions so that you can make sure she understands the material.
Setting Learning Goals Betty is asked a question that she cannot answer for the second time Ive seen this kind of difficultly with teaching other students in the past. You should look for missing links between concepts or links that are in the wrong direction.
11Mentor other forms of help
- On-Demand Help Students select which kind of
helps they need - Pedagogical examples
- What should I teach Betty?
- How can I be sure that Betty learns what I have
taught? - Learning examples
- How do I know that I know enough to teach?
- Domain-content examples
- General What domain content is relevant, chains
of reasoning - Specific I need help on the quiz.
- Help after quiz taken Adaptive
- ICS LBT systems where errors have occurred
in concept map and possible fixes - SRL groups what to read so as to do generate a
more correct map
12Experimental Studies
13Bettys Brain Experimental Studies
- Fifth-grade students teach and learn about river
ecosystems in several 45-min. sessions, and
complete written pre/post tests - Domain River ecosystems interdependence and
balance involving (i) Food Chain, (ii)
Photosynthesis and Respiration, and (iii) Waste
cycle - They later participate in a transfer (PFL) phase
where they learn a new domain (e.g., nitrogen
cycle on land, or global warming). - We have compared several versions of the system
- ICS create a map(no teach) with content
feedback - LBT teach Betty with content feedback
- T-SRL teach Betty with SRL feedback
- M-SRL create a map (no teaching) with SRL
feedback
14Data Analysis
- Performance learning of domain content
- Number of correct concepts links in students
final concept maps - Behaviors sequence of activities
- Key student actions are logged
- Edit map (EM)
- Ask query (AQ)
- Request quiz (RQ)
- Access resources (RA)
- Request explanations (RE/CE)
- Betty could sometimes take (QT) or refuse (QD)
the quiz
15Results Learning Performance
Study 1 ICS, LBT, and T-SRL (56 students)
Condition Mean (SD) Map Scores Mean (SD) Map Scores
Main Phase Transfer Phase
ICS 22.83 (5.3) 22.65 (13.7)
LBT 25.65 (6.5)c 31.81 (12.0)
T-SRL 31.58 (6.6)a,b 32.56 (9.9)a
a T-SRL gt ICS, p lt .05 b T-SRL gt LBT, p lt .05 c
LBT gt ICS, p lt .05.
Study 2 ICS, M-SRL, and T-SRL (83 students)
Condition Mean (SD) Map Scores Mean (SD) Map Scores
Main Phase Transfer Phase
ICS 35.80 (10.5) 36.56 (13.61)
M-SRL 38.41 (8.55) 39.66(16.41)
T-SRL 41.79 (7.37)a 42.97(15.83)
16Behavior Analysis
- Roscoe, et al. (2008) ICS, LBT, and T-SRL in
main study - Map quality was associated with AQ and RE/CE
activities - AQ RE/CE may indicate students attempts to
regulate their own knowledge
17Behavior Analysis using HMMs
- Jeong, et al. (2008) ICS, LBT, and T-SRL in main
and transfer - Used hidden Markov models (HMMs) to model
learning patterns - States hidden, output observer
- Three patterns related to SRL differed by
condition - Map Building EM-RA-RQ
- Map Probing AQ-RA
- Map Tracing AQ-RE-CE
- Interpreted models on right
18Behavior Analysis with HMMs
- Stationary probabilities show the likelihood of
exhibiting a given state
State ICS LBT T-SRL Transfer ICS Transfer LBT Transfer T-SRL
Map Building 0.72 0.66 0.42 0.73 0.73 0.68
Map Probing 0.24 0.30 0.47 0.25 0.25 0.27
Map Tracing 0.04 0.04 0.11 0.02 0.02 0.05
19Pre-Post Test Analysis
- Detailed analyses of students written responses
to examine learning of five river ecosystem
principles - balance, interdependence, microscopic entities,
photosynthesis and cellular respiration,
pollution - Learning about microscopic entities (e.g.,
oxygen, bacteria, and macroinvertebrates) was
strongest - Perhaps, because concept map representations make
normally invisible concepts explicit.
20Current and Future Work
- Adaptive Tutoring through Interactive
metacognition - Betty emulates aspects of self-regulated learner
- Mentor provides additional metacognitive support
to remind students of important cognitive
learning tasks and to help organize these tasks - Further study of self vs self-other monitoring
- Mentor SRL versus Betty SRL
- Increased dose of self-other monitoring
- Front-of-class (FOC) Betty
- Moving TA system into classroom strong links to
science curriculum - Adaptive teaching by the classroom teacher(s)
- Learning science
- From concepts and their relations to causal
reasoning about chain of events
(interdependence) - Aggregate Processes and Balance
http//www.teachableagents.org