d12n d12n - 1 month ago 8
R Question

LDA with topicmodels, how can I see which topics different documents belong to?

I am using LDA from the topicmodels package, and I have run it on about 30.000 documents, acquired 30 topics, and got the top 10 words for the topics, they look very good. But I would like to see which documents belong to which topic with the highest probability, how can I do that?

myCorpus <- Corpus(VectorSource(userbios$bio))
docs <- userbios$twitter_id
myCorpus <- tm_map(myCorpus, tolower)
myCorpus <- tm_map(myCorpus, removePunctuation)
myCorpus <- tm_map(myCorpus, removeNumbers)
removeURL <- function(x) gsub("http[[:alnum:]]*", "", x)
myCorpus <- tm_map(myCorpus, removeURL)
myStopwords <- c("twitter", "tweets", "tweet", "tweeting", "account")

# remove stopwords from corpus
myCorpus <- tm_map(myCorpus, removeWords, stopwords('english'))
myCorpus <- tm_map(myCorpus, removeWords, myStopwords)


# stem words
# require(rJava) # needed for stemming function
# library(Snowball) # also needed for stemming function
# a <- tm_map(myCorpus, stemDocument, language = "english")

myDtm <- DocumentTermMatrix(myCorpus, control = list(wordLengths=c(2,Inf), weighting=weightTf))
myDtm2 <- removeSparseTerms(myDtm, sparse=0.85)
dtm <- myDtm2

library(topicmodels)

rowTotals <- apply(dtm, 1, sum)
dtm2 <- dtm[rowTotals>0]
dim(dtm2)
dtm_LDA <- LDA(dtm2, 30)

Ben Ben
Answer

How about this, using the built-in dataset. This will show you what documents belong to which topic with the highest probability.

library(topicmodels)
data("AssociatedPress", package = "topicmodels")

k <- 5 # set number of topics
# generate model
lda <- LDA(AssociatedPress[1:20,], control = list(alpha = 0.1), k)
# now we have a topic model with 20 docs and five topics

# make a data frame with topics as cols, docs as rows and
# cell values as posterior topic distribution for each document
gammaDF <- as.data.frame(lda@gamma) 
names(gammaDF) <- c(1:k)
# inspect...
gammaDF
              1            2            3            4            5
1  8.979807e-05 8.979807e-05 9.996408e-01 8.979807e-05 8.979807e-05
2  8.714836e-05 8.714836e-05 8.714836e-05 8.714836e-05 9.996514e-01
3  9.261396e-05 9.996295e-01 9.261396e-05 9.261396e-05 9.261396e-05
4  9.995437e-01 1.140774e-04 1.140774e-04 1.140774e-04 1.140774e-04
5  3.573528e-04 3.573528e-04 9.985706e-01 3.573528e-04 3.573528e-04
6  5.610659e-05 5.610659e-05 5.610659e-05 5.610659e-05 9.997756e-01
7  9.994345e-01 1.413820e-04 1.413820e-04 1.413820e-04 1.413820e-04
8  4.286702e-04 4.286702e-04 4.286702e-04 9.982853e-01 4.286702e-04
9  3.319338e-03 3.319338e-03 9.867226e-01 3.319338e-03 3.319338e-03
10 2.034781e-04 2.034781e-04 9.991861e-01 2.034781e-04 2.034781e-04
11 4.810342e-04 9.980759e-01 4.810342e-04 4.810342e-04 4.810342e-04
12 2.651256e-04 9.989395e-01 2.651256e-04 2.651256e-04 2.651256e-04
13 1.430945e-04 1.430945e-04 1.430945e-04 9.994276e-01 1.430945e-04
14 8.402940e-04 8.402940e-04 8.402940e-04 9.966388e-01 8.402940e-04
15 8.404830e-05 9.996638e-01 8.404830e-05 8.404830e-05 8.404830e-05
16 1.903630e-04 9.992385e-01 1.903630e-04 1.903630e-04 1.903630e-04
17 1.297372e-04 1.297372e-04 9.994811e-01 1.297372e-04 1.297372e-04
18 6.906241e-05 6.906241e-05 6.906241e-05 9.997238e-01 6.906241e-05
19 1.242780e-04 1.242780e-04 1.242780e-04 1.242780e-04 9.995029e-01
20 9.997361e-01 6.597684e-05 6.597684e-05 6.597684e-05 6.597684e-05


# Now for each doc, find just the top-ranked topic   
toptopics <- as.data.frame(cbind(document = row.names(gammaDF), 
  topic = apply(gammaDF,1,function(x) names(gammaDF)[which(x==max(x))])))
# inspect...
toptopics   
       document topic
1         1     2
2         2     5
3         3     1
4         4     4
5         5     4
6         6     5
7         7     2
8         8     4
9         9     1
10       10     2
11       11     3
12       12     1
13       13     1
14       14     2
15       15     1
16       16     4
17       17     4
18       18     3
19       19     4
20       20     3

Is that what you want to do?

Hat-tip to this answer: https://stat.ethz.ch/pipermail/r-help/2010-August/247706.html

Comments