Topics can be seen as groups of words that cluster together.
Similar to factor analysis, topic modeling reduces the dimensionality of the feature space (the term-document matrix)
assuming that the latent factors (the topics) will correspond to meaningful latent classes (e.g. issues, frames).
We can use the document-term matrix from the sotu dataset in topicbrowser, which contains the the nouns, names and adjectives from Bush' and Obama's State of the Union addresses.
With this dtm, a topic model can be trained using the LDA function from the topicmodels package.
We specify that we want to use the Gibbs sampler with 200 iterations.
library(topicbrowser)## Loading required package: RColorBrewer
## Loading required package: wordcloud
## Loading required package: Rcpp
## Loading required package: Matrix
## Loading required package: reshape2
## Loading required package: markdown
## Loading required package: base64
##
## Attaching package: 'topicbrowser'
##
## The following object is masked from 'package:knitr':
##
## render_html
library(topicmodels)
library(tm)
data(sotu)
sotu.dtm = weightTf(sotu.dtm)
set.seed(12345)
m = LDA(sotu.dtm, k = 10, method = "Gibbs", control = list(iter = 200))Note that set.seed makes sure that if the function is run again, you get the same results (which would normally not be the case since LDA is non-deterministic).
The resulting LDA model m can be inspected and plotted in a variety of ways,
the simplist of which is listing the most important words per topic using terms:
terms(m, 10)## Topic 1 Topic 2 Topic 3 Topic 4 Topic 5 Topic 6
## [1,] "way" "new" "America" "year" "America" "country"
## [2,] "today" "job" "world" "tax" "energy" "life"
## [3,] "money" "more" "nation" "family" "home" "day"
## [4,] "Security" "business" "people" "last" "goal" "man"
## [5,] "worker" "american" "freedom" "next" "economy" "law"
## [6,] "Social" "tonight" "great" "budget" "power" "woman"
## [7,] "Federal" "company" "free" "deficit" "technology" "citizen"
## [8,] "government" "market" "terror" "more" "clean" "work"
## [9,] "other" "first" "peace" "bill" "oil" "part"
## [10,] "system" "place" "own" "spending" "commitment" "Tonight"
## Topic 7 Topic 8 Topic 9 Topic 10
## [1,] "Americans" "child" "people" "terrorist"
## [2,] "Congress" "school" "american" "Iraq"
## [3,] "health" "good" "time" "States"
## [4,] "care" "education" "America" "United"
## [5,] "reform" "program" "future" "weapon"
## [6,] "plan" "many" "responsibility" "war"
## [7,] "cost" "college" "economy" "security"
## [8,] "insurance" "student" "same" "force"
## [9,] "best" "better" "generation" "enemy"
## [10,] "Medicare" "community" "action" "nuclear"
Although interpreting topics on the top words alone is always iffy, it seems that most of the topics have a distinct meaning. Dot example, topic 10 seems to be about health care, while topic 8 deals with war and security.
The package topicbrowser contains a number of useful functions for plotting topics and listing the top documents per topic.
Let's plot the word ckoud for topic 8:
plot_wordcloud(m, topic_nr = 8)
To make richer visualizations, we combine the LDA model with the original tokens and the metadata into a 'clusterinfo' object:
info = clusterinfo(m, sotu.tokens$lemma, sotu.tokens$aid, words = sotu.tokens$word,
meta = sotu.meta)We can use this to e.g. add topic use over time to the graph:
plot_wordcloud_time(clusterinfo = info, topic_nr = 8, time_interval = "year")
And to create a network visualization rather than a normal word cloud:
plot_semnet(clusterinfo = info, topic_nr = 10)## Loading required package: semnet
## Loading required package: plyr
## Loading required package: zoo
##
## Attaching package: 'zoo'
##
## The following objects are masked from 'package:base':
##
## as.Date, as.Date.numeric
##
## Loading required package: scales
## Loading required package: slam
## Loading required package: igraph
## Note: method with signature 'CsparseMatrix#Matrix#missing#replValue' chosen for function '[<-',
## target signature 'dgCMatrix#nsCMatrix#missing#numeric'.
## "Matrix#nsparseMatrix#missing#replValue" would also be valid
Finally, we can create a 'topic browser' HTML page which contains an overview of all topics. See e.g. the example at rpubs
createTopicBrowser(info)We can also include the 'semantic network' topic representations by including that in the per-topic plot functions:
createTopicBrowser(info, plotfunction.pertopic = c(plot_wordcloud_time, plot_semnet))