Process Mining Student Discourse
Student Epistemics
Tanner Phillips
1/31/2022
Introduction
I’ve included this code in my portfolio not because it’s the most complicated, but because it shows some very basic principles of coding. Functions, efficient looping, and data manipulation. There’s very little tidyverse here, just base R. It’s easy to write this code badly.
The underlying problem is turning a list of items into a frequency matrix based on complex conditions. The data is student chat from a computer-supported collaborative learning environment (i.e., an educational video game). We were attempting to use processes mining to understand the order of types of speech (e.g., question, assertions of fact, social organization). To do this we wanted to get frequency counts of pairs and triplets of types of speach, like a question, followed by statement of fact, followed by another question.
I’m very proud of (a) The efficient speed at which this code ran when applied 10,000 lines of student chat, and (b) figuring out a simple way to visualize the data that allowed us to make interesting inferences about the data. As of February 2022, we are currently in the process of writing this up as a journal article.
Definitions
The definitions of the different codes may help with understanding the ouput:
K-. A question or other query for information. Often in speech our questions are implicit, not explicit.
K+. A direct knowledge claim or assertion.
Reply. A reply to previous comment. Would be meaningless outside of context.
Reply - Knowledge. A reply that includes new knowledge.
Reply - Hedge. A reply that “hedges” what is being said (e.g. I’m not sure, but… etc.)
Social Organization. Attempts to organize from introductions like “hello” to more explicit organizaton like “what should our next step be.
Other. Anything else. Often spam or off-topic
Code
library(tidyverse)## -- Attaching packages --------------------------------------- tidyverse 1.3.1 --## v ggplot2 3.3.5 v purrr 0.3.4
## v tibble 3.1.4 v dplyr 1.0.7
## v tidyr 1.1.3 v stringr 1.4.0
## v readr 2.0.1 v forcats 0.5.1## -- Conflicts ------------------------------------------ tidyverse_conflicts() --
## x dplyr::filter() masks stats::filter()
## x dplyr::lag() masks stats::lag()library(RColorBrewer)
#Support constant is used to determine the % Cutoff for patterns in student speech. i.e. if less that support % of speech patterns fall into this category, we don't carry them forward.
support<-.05Custom Functions
#Custom Function to calculate marginal percentages for rows or columns of a matrix
margin_prop<-function(x){
s<-sum(x)
x/s
}
# Custom function used to select locations in the data where a certain sequence of 2 or 3 codes is present
epis_pattern<-function(vector,pattern1,pattern2,pattern3=NA){
locations<-c()
if(is.na(pattern3)){
for(i in 1:(length(vector)-1)){
if(vector[i]==pattern1 & vector[i+1]==pattern2){
locations<-c(locations,i)
}
}
}else{
for(i in 1:(length(vector)-2)){
if(vector[i]==pattern1 & vector[i+1]==pattern2 & vector[i+2]==pattern3){
locations<-c(locations,i)
}
}
}
return(locations+1)
}Load and Clean
codes<-read.csv("EcoJourney_DiscourseData_All_v2.csv")
codes<-codes[,1:7]
names(codes)[1]<-"GroupID"
#Split out the epistemic column. This is just for convenience.
epis<-codes$Epistemics
Wizard.indecies<-grepl("w",codes$UserID,ignore.case = T)
Wizard.indecies.IDS<-which(grepl("w",codes$UserID,ignore.case = T))Analysis 1: Epistemic Pairs
###Create epistemic pairs matrix
freq1<-matrix(0,nrow=length(unique(epis)),ncol=length(unique(epis)))
wizard_freq1<-matrix(0,nrow=length(unique(epis)),ncol=length(unique(epis)))
###Name rows and columns. We'll use these to select cells in the matrices
rownames(freq1)<-unique(epis)
colnames(freq1)<-unique(epis)
rownames(wizard_freq1)<-unique(epis)
colnames(wizard_freq1)<-unique(epis)
###Comb through coded text and find couplets of speech types and put frequencies in matrix.
for(i in 1:(length(epis)-1)){
freq1[epis[i],epis[i+1]]<-(freq1[epis[i],epis[i+1]]+1)
if(i %in% Wizard.indecies.IDS){
wizard_freq1[epis[i],epis[i+1]]<-(wizard_freq1[epis[i],epis[i+1]]+1)
}
}
###Transform into frequencies.
freq1_prob<-t(apply(freq1,MARGIN=1,margin_prop))
rownames(freq1_prob)<-names(freq1_prob)
matrix_rows = sum(freq1_prob > support)
support1<-matrix(0,nrow = matrix_rows,ncol=2)
###Select all supported discourse pairs to carry forward into triplet analysis.
k = 1
for(i in 1:nrow(freq1_prob)){
for(j in 1:ncol(freq1_prob)){
if(freq1_prob[i,j]>support){
support1[k,]<-c(rownames(freq1)[i],colnames(freq1)[j])
k = k + 1
}
}
}
freq1## Other Social organization K- K+ Reply Reply-Hedge
## Other 1103 90 146 151 213 23
## Social organization 60 156 104 92 144 24
## K- 133 47 254 147 581 131
## K+ 127 65 208 310 383 39
## Reply 259 184 534 393 1440 123
## Reply-Hedge 29 18 91 35 158 81
## Reply-Knowledge 47 38 167 52 156 52
## Reply-Knowledge
## Other 32
## Social organization 18
## K- 211
## K+ 48
## Reply 142
## Reply-Hedge 61
## Reply-Knowledge 214Analysis 2: Epistemic Triplets
###Initiate matrices
freq2<-matrix(0,nrow=nrow(support1),ncol=nrow(freq1))
wizard_freq2<-matrix(0,nrow=nrow(support1),ncol=nrow(freq1))
###name rows and columns
rownames(freq2)<-paste(support1[,1],support1[,2],sep = "->")
colnames(freq2)<-colnames(freq1)
rownames(wizard_freq2)<-paste(support1[,1],support1[,2],sep = "->")
colnames(wizard_freq2)<-colnames(freq1)
###Grab Epistemic Triplets. Essentialy same as for couplets.
for(i in 1:(length(epis)-3)){
pattern = paste(epis[i],epis[i+1],sep = "->")
if(pattern %in% rownames(freq2)){
freq2[pattern,epis[i+2]]<-freq2[pattern,epis[i+2]]+1
if(i %in% Wizard.indecies.IDS){
wizard_freq2[pattern,epis[i+2]]<- wizard_freq2[pattern,epis[i+2]]+1
}
}
}
###Save percentage frequencies by row
freq2_prob<-as.data.frame(t(apply(freq2,MARGIN=1,margin_prop)))
colnames(freq2_prob)<-colnames(freq2)
head(freq2)## Other Social organization K- K+ Reply Reply-Hedge
## Other->Other 828 47 61 83 69 8
## Other->Social organization 18 30 10 15 15 2
## Other->K- 37 3 27 13 43 6
## Other->K+ 46 8 23 36 32 3
## Other->Reply 51 5 30 22 90 7
## Social organization->Other 31 8 9 6 5 0
## Reply-Knowledge
## Other->Other 5
## Other->Social organization 0
## Other->K- 17
## Other->K+ 3
## Other->Reply 8
## Social organization->Other 1Results
Unpacking these results is the topic of the paper we are currently writing and would make this document a bit unruly. See the “Student Discourse Results” powerpoint in my public github repo if you’re interest in a quick overview.
###Full Results
heatmap(freq1,
main="Chat Pair Frequencies",
margins = c(12,12),
ylab="First Chat",
xlab="Second Chat")
heatmap(as.matrix(freq2_prob),
margins=c(12,12),
main="Chat Triplet Frequencies",
ylab="First & Second Chat",
xlab="Third Chat")
###Wizard Results
heatmap(as.matrix(wizard_freq1),
main="Chat Pairs Initiated by Instructor",
margins=c(12,12),
ylab="First Chat",
xlab="Second Chat")
heatmap(as.matrix(wizard_freq2),
main="Chat triplets Initiated by Instructor",
xlab="Third Chat",
margins=c(12,12))
####Select Lines for qualitative review
lines<-epis_pattern(epis,"Reply-Knowledge","K-","K-")
lines## [1] 247 1522 1993 2309 2412 4980 5226 5663 5677 5685 5732 6764 7496 7515 7528
## [16] 7694 8389 8846 8921