Search code examples
rnlptmword-cloudsnowball

Making a wordcloud, but with combined words?

I am trying to make a word cloud of publications keywords. for example: Educational data mining; collaborative learning; computer science...etc

My current code is as the following:

KeywordsCorpus <- Corpus(VectorSource(subset(Words$Author.Keywords, Words$Year==2012)))
KeywordsCorpus <- tm_map(KeywordsCorpus, removePunctuation)
KeywordsCorpus <- tm_map(KeywordsCorpus, removeNumbers)

# added tolower
KeywordsCorpus <- tm_map(KeywordsCorpus, tolower)
KeywordsCorpus <- tm_map(KeywordsCorpus, removeWords, stopwords("english"))

# moved stripWhitespace
KeywordsCorpus <- tm_map(KeywordsCorpus, stripWhitespace)  

KeywordsCorpus <- tm_map(KeywordsCorpus, PlainTextDocument)

dtm4 <- TermDocumentMatrix(KeywordsCorpus)
m4 <- as.matrix(dtm4)
v4 <- sort(rowSums(m4),decreasing=TRUE)
d4 <- data.frame(word = names(v4),freq=v4)

However, With this code, it is splitting each word by itself, But what I need is to have a combined words/phrases. For instance: Educational Data Mining is 1 phrase that I need to show instead of what is happening: "Educational" "Data" "Mining". Is there a way to keep each compound of words together? the semi-colon might help as a separator.

Thanks.

Solution

  • Here's a solution using a different text package, that allows you to form multi-word expressions from either statistically detected collocations, or just by forming all bi-grams. The package is called quanteda.

    library(quanteda)
packageVersion("quanteda")
## [1] ‘0.9.5.14’

    First, the method for detecting the top 1,500 bigram collocations, and replacing these collocations in the texts with their single-token versions (concatenated by the "_" character). Here I am using the package's built-in corpus of the US presidential inaugural address texts.

    ### for just the top 1500 collocations
# detect the collocations
colls <- collocations(inaugCorpus, n = 1500, size = 2)

# remove collocations containing stopwords
colls <- removeFeatures(colls, stopwords("SMART"))
## Removed 1,224 (81.6%) of 1,500 collocations containing one of 570 stopwords.

# replace the phrases with single-token versions
inaugCorpusColl2 <- phrasetotoken(inaugCorpus, colls)

# create the document-feature matrix
inaugColl2dfm <- dfm(inaugCorpusColl2, ignoredFeatures = stopwords("SMART"))
## Creating a dfm from a corpus ...
## ... lowercasing
## ... tokenizing
## ... indexing documents: 57 documents
## ... indexing features: 9,741 feature types
## ... removed 430 features, from 570 supplied (glob) feature types
## ... complete. 
## ... created a 57 x 9311 sparse dfm
## Elapsed time: 0.163 seconds.

# plot the wordcloud
set.seed(1000)
png("~/Desktop/wcloud1.png", width = 800, height = 800)
plot(inaugColl2dfm["2013-Obama", ], min.freq = 2, random.order = FALSE, 
     colors = sample(colors()[2:128]))
dev.off()

    This results in the following plot. Note the collocations, such as "generation's_task" and "fellow_americans".

    wcloud1.png

    The version formed with all bigrams is easier, but results in a huge number of low frequency bigram features. For the word cloud, I selected a larger set of texts, not just the 2013 Obama address.

    ### version with all bi-grams
inaugbigramsDfm <- dfm(inaugCorpusColl2, ngrams = 2, ignoredFeatures = stopwords("SMART"))
## Creating a dfm from a corpus ...
## ... lowercasing
## ... tokenizing
## ... indexing documents: 57 documents
## ... removed 54,200 features, from 570 supplied (glob) feature types
## ... indexing features: 64,108 feature types
## ... created a 57 x 9908 sparse dfm
## ... complete. 
## Elapsed time: 3.254 seconds.

# plot the bigram wordcloud - more texts because for a single speech, 
# almost none occur more than once
png("~/Desktop/wcloud2.png", width = 800, height = 800)
plot(inaugbigramsDfm[40:57, ], min.freq = 2, random.order = FALSE, 
     colors = sample(colors()[2:128]))
dev.off()

    This produces:

    wcloud2.png