How to create a table by restructuring a MALLET output file?
https://stackoverflow.com/questions/8058402
italiano
english
français
española
中国
日本の
العربية
Deutsch
한국어
Português
RussianQuestion
I'm using MALLET for topic analysis which is outputting results in text files ("topics.txt") of several thousand rows and a hundred or so rows where each row consists of tab-separated variables like this:
Num1 text1 topic1 proportion1 topic2 proportion2 topic3 proportion3, etc.
Num2 text2 topic1 proportion1 topic2 proportion2 topic3 proportion3, etc.
Num3 text3 topic1 proportion1 topic2 proportion2 topic3 proportion3, etc.
Here's a snippet of the actual data:
> dat[1:5,1:10]
V1 V2 V3 V4 V5 V6 V7 V8 V9 V10
1 0 10.txt 27 0.4560785 23 0.3040853 20 0.1315621 21 0.03632624
2 1 1001.txt 20 0.2660085 12 0.2099153 8 0.1699586 13 0.16922928
3 2 1002.txt 16 0.3341721 2 0.1747023 10 0.1360454 12 0.07507119
4 3 1003.txt 12 0.5366148 8 0.2255179 18 0.1388561 0 0.01867091
5 4 1005.txt 16 0.2363206 0 0.2214441 24 0.1914769 7 0.17760521
I'm trying to use R to convert this output into a data table where the topics are column headers and each topic contains the values of the variable 'proportion' directly to the right hand side of each variable 'topic', for each value of 'text'. Like this:
topic1 topic2 topic3
text1 proportion1 proportion2 proportion3
text2 proportion1 proportion2 proportion3
or with the data snippet above, like so:
0 2 7 8 10 12 13 16 18 20 21 23 24 27
10.txt 0 0 0 0 0 0 0 0 0 0.1315621 0.03632624 0.3040853 0 0.4560785
1001.txt 0 0 0 0.1699586 0 0.2099153 0.1692292 0 0 0.2660085 0 0 0 0
1002.txt 0 0.1747023 0 0 0.1360454 0.0750711 0 0.3341721 0 0 0 0 0 0
1003.txt 0.0186709 0 0 0.2255179 0 0.5366148 0 0 0.138856 0 0 0 0 0
1005.txt 0.2214441 0 0.1776052 0 0 0 0 0.2363206 0 0 0 0 0.1914769 0
This is the R code I've got to do the job, sent from a friend, but it doesn't work for me (and I don't know enough about it to fix it myself):
##########################################
dat<-read.table("topics.txt", header=F, sep="\t")
datnames<-subset(dat, select=2)
dat2<-subset(dat, select=3:length(dat))
y <- data.frame(topic=character(0),proportion=character(0),text=character(0))
for(i in seq(1, length(dat2), 2)){
z<-i+1
x<-dat2[,i:z]
x<-cbind(x, datnames)
colnames(x)<-c("topic","proportion", "text")
y<-rbind(y, x)
}
# Right at this step at the end of the block
# I get this message that may indicate the problem:
# Error in c(in c("topic", "proportion", "text") : unused argument(s) ("text")
y[is.na(y)] <- 0
xdat<-xtabs(proportion ~ text+topic, data=y)
write.table(xdat, file="topicMatrix.txt", sep="\t", eol = "\n", quote=TRUE, col.names=TRUE, row.names=TRUE)
##########################################
I'd be most grateful for any suggestions on how I can get this code working. My problem may be related to this one and possibly this one also, but I don't yet have the skills to make immediate use of the answers to those questions.
Solution
Here is one approach to your problem
dat <-read.table(as.is = TRUE, header = FALSE, textConnection(
"Num1 text1 topic1 proportion1 topic2 proportion2 topic3 proportion3
Num2 text2 topic1 proportion1 topic2 proportion2 topic3 proportion3
Num3 text3 topic1 proportion1 topic2 proportion2 topic3 proportion3"))
NTOPICS = 3
nam <- c('num', 'text',
paste(c('topic', 'proportion'), rep(1:NTOPICS, each = 2), sep = ""))
dat_l <- reshape(setNames(dat, nam), varying = 3:length(nam), direction = 'long',
sep = "")
reshape2::dcast(dat_l, num + text ~ topic, value_var = 'proportion')
num text topic1 topic2 topic3
1 Num1 text1 proportion1 proportion2 proportion3
2 Num2 text2 proportion1 proportion2 proportion3
3 Num3 text3 proportion1 proportion2 proportion3
EDIT. This will work irrespective of whether the proportions are text or numbers. You can also modify NTOPICS to suit the number of topics you have
OTHER TIPS
You can get this into a long format but to go further required real data. EDITED after data offered. Still not sure about the overall structure of what is coming from MALLET, but at least the R functions are demonstrated. This approach has the "feature" that proportions are summed if there are overlapping topics. Depending on the data layout that may be an advantage or not.
dat <-read.table(textConnection(" V1 V2 V3 V4 V5 V6 V7 V8 V9 V10
1 0 10.txt 27 0.4560785 23 0.3040853 20 0.1315621 21 0.03632624
2 1 1001.txt 20 0.2660085 12 0.2099153 8 0.1699586 13 0.16922928
3 2 1002.txt 16 0.3341721 2 0.1747023 10 0.1360454 12 0.07507119
4 3 1003.txt 12 0.5366148 8 0.2255179 18 0.1388561 0 0.01867091
5 4 1005.txt 16 0.2363206 0 0.2214441 24 0.1914769 7 0.17760521
"),
header=TRUE)
ldat <- reshape(dat, idvar=1:2, varying=list(topics=c("V3", "V5", "V7", "V9"),
props=c("V4", "V6", "V8", "V10")),
direction="long")
####------------------####
> ldat
V1 V2 time V3 V4
0.10.txt.1 0 10.txt 1 27 0.45607850
1.1001.txt.1 1 1001.txt 1 20 0.26600850
2.1002.txt.1 2 1002.txt 1 16 0.33417210
3.1003.txt.1 3 1003.txt 1 12 0.53661480
4.1005.txt.1 4 1005.txt 1 16 0.23632060
0.10.txt.2 0 10.txt 2 23 0.30408530
1.1001.txt.2 1 1001.txt 2 12 0.20991530
2.1002.txt.2 2 1002.txt 2 2 0.17470230
3.1003.txt.2 3 1003.txt 2 8 0.22551790
4.1005.txt.2 4 1005.txt 2 0 0.22144410
0.10.txt.3 0 10.txt 3 20 0.13156210
1.1001.txt.3 1 1001.txt 3 8 0.16995860
2.1002.txt.3 2 1002.txt 3 10 0.13604540
3.1003.txt.3 3 1003.txt 3 18 0.13885610
4.1005.txt.3 4 1005.txt 3 24 0.19147690
0.10.txt.4 0 10.txt 4 21 0.03632624
1.1001.txt.4 1 1001.txt 4 13 0.16922928
2.1002.txt.4 2 1002.txt 4 12 0.07507119
3.1003.txt.4 3 1003.txt 4 0 0.01867091
4.1005.txt.4 4 1005.txt 4 7 0.17760521
Now can show you how to use xtabs() since those "proportions" are "numeric". Something like this may eventually be what you want. I was surprised that the topics were also integers but perhaps there is a mapping from topic numbers to topic names?:
> xtabs(V4 ~ V3 + V2, data=ldat)
V2
V3 10.txt 1001.txt 1002.txt 1003.txt 1005.txt
0 0.00000000 0.00000000 0.00000000 0.01867091 0.22144410
2 0.00000000 0.00000000 0.17470230 0.00000000 0.00000000
7 0.00000000 0.00000000 0.00000000 0.00000000 0.17760521
8 0.00000000 0.16995860 0.00000000 0.22551790 0.00000000
10 0.00000000 0.00000000 0.13604540 0.00000000 0.00000000
12 0.00000000 0.20991530 0.07507119 0.53661480 0.00000000
13 0.00000000 0.16922928 0.00000000 0.00000000 0.00000000
16 0.00000000 0.00000000 0.33417210 0.00000000 0.23632060
18 0.00000000 0.00000000 0.00000000 0.13885610 0.00000000
20 0.13156210 0.26600850 0.00000000 0.00000000 0.00000000
21 0.03632624 0.00000000 0.00000000 0.00000000 0.00000000
23 0.30408530 0.00000000 0.00000000 0.00000000 0.00000000
24 0.00000000 0.00000000 0.00000000 0.00000000 0.19147690
27 0.45607850 0.00000000 0.00000000 0.00000000 0.00000000
Returning to this problem, I've found that the reshape function is far too demanding on memory, so I use a data.table method instead. A few more steps, but a huge amount faster and substantially less memory intensive.
dat <- read.table(text = "V1 V2 V3 V4 V5 V6 V7 V8 V9 V10
1 0 10.txt 27 0.4560785 23 0.3040853 20 0.1315621 21 0.03632624
2 1 1001.txt 20 0.2660085 12 0.2099153 8 0.1699586 13 0.16922928
3 2 1002.txt 16 0.3341721 2 0.1747023 10 0.1360454 12 0.07507119
4 3 1003.txt 12 0.5366148 8 0.2255179 18 0.1388561 0 0.01867091
5 4 1005.txt 16 0.2363206 0 0.2214441 24 0.1914769 7 0.17760521")
dat$V11 <- rep(NA, 5) # my real data has this extra unwanted col
dat <- data.table(dat)
# get document number
docnum <- dat$V1
# get text number
txt <- dat$V2
# remove doc num and text num so we just have topic and props
dat1 <- dat[ ,c("V1","V2", paste0("V", ncol(dat))) := NULL]
# get topic numbers
n <- ncol(dat1)
tops <- apply(dat1, 1, function(i) i[seq(1, n, 2)])
# get props
props <- apply(dat1, 1, function(i) i[seq(2, n, 2)])
# put topics and props together
tp <- lapply(1:ncol(tops), function(i) data.frame(tops[,i], props[,i]))
names(tp) <- txt
# make into long table
dt <- data.table::rbindlist(tp)
dt$doc <- unlist(lapply(txt, function(i) rep(i, ncol(dat1)/2)))
dt$docnum <- unlist(lapply(docnum, function(i) rep(i, ncol(dat1)/2)))
# reshape to wide
library(data.table)
setkey(dt, tops...i., doc)
out <- dt[CJ(unique(tops...i.), unique(doc))][, as.list(props...i.), by=tops...i.]
setnames(out, c("topic", as.character(txt)))
# transpose to have table of docs (rows) and columns (topics)
tout <- data.table(t(out))
setnames(tout, unname(as.character(tout[1,])))
tout <- tout[-1,]
row.names(tout) <- txt
# replace NA with zero
tout[is.na(tout)] <- 0
And here's the output, docs as rows, topics as columns, doc names are in the rownames, which are not printed, but available for later use.
tout
0 2 7 8 10 12 13 16 18
1: 0.00000000 0.0000000 0.0000000 0.0000000 0.0000000 0.00000000 0.0000000 0.0000000 0.0000000
2: 0.00000000 0.0000000 0.0000000 0.1699586 0.0000000 0.20991530 0.1692293 0.0000000 0.0000000
3: 0.00000000 0.1747023 0.0000000 0.0000000 0.1360454 0.07507119 0.0000000 0.3341721 0.0000000
4: 0.01867091 0.0000000 0.0000000 0.2255179 0.0000000 0.53661480 0.0000000 0.0000000 0.1388561
5: 0.22144410 0.0000000 0.1776052 0.0000000 0.0000000 0.00000000 0.0000000 0.2363206 0.0000000
20 21 23 24 27
1: 0.1315621 0.03632624 0.3040853 0.0000000 0.4560785
2: 0.2660085 0.00000000 0.0000000 0.0000000 0.0000000
3: 0.0000000 0.00000000 0.0000000 0.0000000 0.0000000
4: 0.0000000 0.00000000 0.0000000 0.0000000 0.0000000
5: 0.0000000 0.00000000 0.0000000 0.1914769 0.0000000
