Educational data mining in a

computer tutor that listens

Joseph E. Beck
Acknowledgements: NSF, Heinz

1
 Take away point
Computer tutors provide gold mine Standardized tests
of fine-grained interaction data

Creates the ability to assess students and

improve capabilities of computer tutors

2
 What is educational data mining?
• Using data to learn about students and
instruction
– E.g. predict student behavior, assess
students, evaluate the tutor’s teaching, etc.
• Motivation: computer tutors provide large
samples of fine-grained, longitudinal data
that are a powerful (unique?) source of
knowledge to improve educational
outcomes

3
 Difference between educational
and standard data mining
• Data collected with purpose in mind
– Have control over schema
• Describe more interactive phenomena
• Generally smaller datasets

4
Project LISTEN’s Reading Tutor
 Data we’ve collected
• Record several items in database
– Student’s speech (as recognized by ASR)
– Student’s help requests
– Tutor’s teaching actions
– (among other things)
• Scale of DB from 2002-2003 school year
– 456 students
– 423,149 student clicks for help
– 4.1 million words heard by speech recognizer
6
 Outline
 Predicting student behavior
• E.g. will the student click for help on this
word?
 Inferring student’s skills
• E.g. does the student know “ch” can make a K
sound (e.g. “chaos”)?
 Future work

7
 Predicting student help requests

8
 Why predict help requests?
• Goal is to understand the student
– In two senses
• A good model of the student should be able to
predict future actions (e.g. outcome measure)
• Help requests provide window into student’s
reading proficiency (e.g. source of knowledge)
• Non-speech events are less noisy
• Applications of help requests
– Avoid overly complex material
– Provide help ahead of time
9
Learning curves in students’ help
requests (with Peng Jia)
Mean help request rate .4

.3

.2
Reading level
Grade 1

.1 Grade 2

Grade 3

0.0 Grade 4
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15

10
Number of previous encounters
 Region of focus

• “When students need help” .4

– 1st and 2nd grade ability

– 1-6 prior word encounters .3
• Selected data

Mean help request rate

– 53 students
– 175,961 words .2
Reading level
– 29,278 help requests Grade 1

• # of cases per student: .1 Grade 2

1392 - 7783 Grade 3

• Help rate excluding 0.0 Grade 4

common words: 0.54%– 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15

54% Number of previous encounters

– A few novice readers
requested substantial 11
amounts of help
 How to predict help requests

• Approach: treat as classifier learning problem

– Inputs: features about the word and the student
– Output: whether the student will ask for help

• Need to decide:
– Features describing word and student
– What data to use to train model

12
 Abbreviated example of features
(20 features were used)
Student on Student
Word
this word overall Requested
Seen Helped Help help?
Length Frequency Grade
before? before? rate

6 1189 Yes No 0.5 1 Yes

11 22255 No No 0.5 1 Yes

3 826 Yes Yes 0.1 3 No

5 1537 No No 0.05 2 No
.
.
.
 Grouped student prediction
• Predict whether student will request help
by using other students’ data
• Leave one student out cross validation:
– Training data: randomly select 25% of all
other students and pool their data together
• (Using all data crashed the machine.)
– Testing data: student’s data

14
Grouped model prediction results
• Used J48 (version of C4.5) and NBC
• Evaluation criteria: weighted accuracy
– Weigh cases where student asked for help 5 times
more heavily
– Not providing help when needed worse than extra help
• Performance (averaged per-student)
– J48: 71%
– NBC: 75%
• How to (possibly) do better?
– Build individualized models for each student
15
 Building individual models
training testing

beginning All data ordered by time

testing
In the
middle Training

• Incrementally construct models as data are seen

• Same features as grouped student prediction
• Performance (averaged per-student)
– J48: 81%
– NBC: 75%
– Better to use data about individuals than population
• Obvious extension: combine grouped and individual
modeling approaches 16
 Using subword properties to help
predict help requests
(with June Sison)
• If student is predicted to need help on
“chord,” he would probably need help on
“chords” as well
– Word roots?
– But what about “chaos?” “chemical?”
• CH/K/ is common across items
• Model lettersound mappings in words
– Called graphemephoneme (gp) mappings
17
 Which gp mappings to use?
• Chemical
– CHK
– EEH
– MM
– IAH
– CK
– AAH
– LL

18
 Which gp mappings to use?
• Chemical
– CHK
– EEH
– MM
– IAH
– CK
– AAH
– LL
• First and last parts of a word are most important for
children’s decoding (Perfetti)
– And adults’ decoding (recent email message floating around)
>Aoccdrnig to rsereach at an Elingsh uinervtisy, it deosn't mttaer in waht
>oredr the ltteers in a wrod are, the olny iprmoetnt tihng is taht frist and
>lsat ltteres are in the rghit pclae…Tihs is bcuseae we do not raed ervey
>lteter by istlef, but the wrod as a wlohe.
19
 Features describing a gp
• P(g): How common is this grapheme?
• P(p|g): How likely is it to generate this
sound given the letters?
• Compute above two features for
– First gp in a word
– Rarest gp in a word
– Average of all gp in a word
• Add to classifier’s set of features
20
 Results
• Used individual models with J48
• Improved accuracy by 0.7% absolute (P=0.013)
over not using gp features
– However, already using many features about student
– Suggests students are sensitive to gp properties
• Can we do better?
– These gp properties are static
– Only describe words, not students
– Perhaps modeling a student’s skills would work
better?
• Infer what is in student’s head rather than just predict actions
21
 Outline
 Predicting student behavior
• E.g. will the student click for help on this
word?
 Inferring student’s skills
• E.g. does the student know “ch” can make a K
sound (e.g. “chaos”)?
 Future work

22
  Automated assessment
(with Peng Jia and June Sison)
• We gather lots of data; use it to assess students
– “Knowing What Students Know” provides metaphor
• Why perform automated assessment?
– Drawbacks of paper tests:
• Expensive
• Lack of ongoing results
• Costly to report to teachers and computer tutors
• Problem: our data are (literally) noisy
– But we have a lot of it: students attempt over 300 words
per day

23
Converting speech input to usable data
Speech input (Sphinx)

Align text I’LL HAVE TO MOP UP MUTTERED DENNIS…

(Multimatch) “I'll have to mop it all up,” muttered Dennis…

Assess subword knowledge

24
 Assessing subword knowledge
• Interested in student proficiency in
individual gp mappings
– Maintain knowledge estimate, P(knows), for
each mapping
• “Hidden subskill problem” (latent variable)
– Cannot assess directly
• Credit/blame first and last gp of every
word attempted
– But how?
25
 What is knowledge tracing?
(Corbett et al.)
Unlearned p(T) Learned
State State
p(L0)
p(G) 1-p(S)

correct correct
Two Learning Parameters
p(L0) Probability the rule is in the learned state at time 0 (prior to the first
opportunity to apply the rule in problem solving).
p(T) Probability the rule will make the transition from the unlearned state to the
learned state at each opportunity to apply the rule
Two Performance Parameters
p(G) Probability the student will guess correctly if the rule is in the unlearned state
p(S) Probability the student will slip (make a mistake) if the rule is in the learned 26
state
 Modifying knowledge tracing
• Problem: noisy speech recognition
• Solution: broaden notion of slip and guess
– P(slip) = chance a skilled student makes a mistake +
chance ASR fails to hear correct reading
– P(guess) = chance a novice pronounces word correctly +
chance ASR incorrectly credits student
• Very different semantics of slip/guess
• Knowledge tracing equations unchanged
• Estimate slip/guess from students working
with Reading Tutor
27
 Applying knowledge tracing
• E.g. Student reads “Dennis” correctly
update DD, SS
• Assume student had P(knows) = 0.1 for both

• Update P(knows DD) • Update P(knows SS)

– P(guess DD) = 0.81 – P(guess SS) = 0.80
– P(slip DD) = 0.13 – P(slip SS) = 0.12
– New P(knows DD) = 0.107 – New P(knows SS) = 0.109

• Slow updates
– A good thing
28
 Evaluation of gp mappings
• Data from 2002-2003
• N=259 (1st through 4th graders)
• Goal: predict performance on fluency posttest
– Standardized test is scored by humans
– (Not our final goal)
• Construct 2 linear models for all students
– Mean P(knows) for all gp
• Fluency posttest = 133.3 * mean – 42.8
– Pretest paper-test score
• Fluency posttest = 0.809 * fluency pretest + 20.5 29
 Results
• All results are leave-one-out cross validation
– Correlation of 0.862 for P(knows) for all gp
– Correlation of 0.808 for pretests
• Look at within-grade correlation
– Reduce heterogeneity
– E.g. shoe size and spelling ability
Grade Mean Pretest N
1 0.878 0.561 115
2 0.808 0.881 80
3 0.813 0.883 42
4 0.859 0.898 22
Average 0.840 0.806 30
Using mean of P(knows) to predict
GORT posttest
100

90

80
Actual GORT postest score

70

60

50

40

30

20

10

0
0 10 20 30 40 50 60 70 80 90 100
Predicted GORT posttest score

31
 Outline
 Predicting student behavior
• E.g. will the student click for help on this
word?
 Inferring student’s skills
• E.g. does the student know “ch” can make a K
sound (e.g. “chaos”)?
 Future work

32
 Near-term goals
• Construct more general tools
– Crosstabs
– View a student’s growth in reading

• Automated assessment
– Validate gp mappings
– Model latent variables
– Improve ASR

33
 Model of student knowledge
Speech

GP knowledge … DD DT DDZ … ZZZ

(371 items!)

Reading proficiency
34
 Model of student knowledge:
adding latent variables
Speech

GP knowledge … DD DT DDZ … ZZZ

“Higher level”
knowledge
e.g. short vowels,
rare use, etc.

Reading proficiency
35
 Improving ASR
• Cannot listen for all mistakes
• Bias ASR based on student proficiencies
• E.g. student encounters “thugs”
– thth, uah, gg, sz
P(say “Thugs”) = 0.90
P(say “Tugs”) = 0.02
 Improving ASR
• Cannot listen for all mistakes
• Bias ASR based on student proficiencies
• E.g. student encounters “thugs”
– thth, uah, gg, sz
P(say “Thugs”) = 0.90
P(say “Tugs”) = 0.02

37
 Improving ASR
• Cannot listen for all mistakes
• Bias ASR based on student proficiencies
• E.g. student encounters “thugs”
– thth, uah, gg, sz
P(say “Thugs”) = 0.90
P(say “Tugs”) = 0.02
Knows theta Doesn’t know theta

P(say “Thugs”) = 0.95 P(say “Thugs”) = 0.40

P(say “Tugs”) = 0.01 P(say “Tugs”) = 0.40
 Longer-term goals
• Improving ASR good goal due to ability to
evaluate changes offline
• However, would like to improve educational
outcomes
– Problem: harder to evaluate learning since human
trials are expensive
– Solution: construct a simulation of the student and
tutor and use reinforcement learning (RL)
– Approach used in my dissertation work (at UMass)
on ADVISOR in the AnimalWatch system

39
 ADVISOR overview
Data from prior Teaching
users of tutor goal

Teaching action
“try again”

Predict student Pedagogical

behavior in state s Agent

Result “correct answer,

took 15 sec.”

Teaching policy
40
 Why is applying RL harder in the
Reading Tutor than in AnimalWatch?
• AnimalWatch • Reading Tutor
– I designed – Built by others
– Started with student – Still building student
model model
– Domain was math, – Domain is reading,
easy to measure hard to measure
outcomes outcomes
– Built from ground up – Greater variety of
with ADVISOR in mind points to intervene

41
 Conclusions
• Can assess students despite noisy data
• For predicting student behavior, data are plentiful
• Can examine models and features
• For predicting student test scores, data are scarce
• Restricted to simple models and need good features
• Educational data mining offers many opportunities
to improve efficacy of teaching
• Big data is a “secret weapon” but…
– We still don’t have enough to do everything we want

42