Make less sparsity when selecting features in text mining - r

Question

I try to create a sentiment analysis of customer reviews using R .
Here the R code :
train.index <- 1:50000
test.index <- 50001:nrow(Data_clean)
#clean up
# remove grammar/punctuation
Data_clean$Raison.Reco.clean <- tolower(gsub('[[:punct:]0-9]', '', Data_clean$Raison.Reco))
train <- Data_clean[train.index, ]
test <- Data_clean[test.index, ]
word.freq <- function(document.vector, sparsity = .999)
{
temp.tf <- document.vector %>% tokens(ngrams = 1:3) %>% # generate tokens
dfm # generate dfm
freq.df <- colSums(temp.tf)
freq.df <- data.frame(word = names(freq.df), freq = freq.df)
rownames(freq.df) <- NULL
return(freq.df)
}
word.freq.pos <- word.freq(train$Raison.Reco.clean[train$Note.Reco == "10"])
word.freq.neg <- word.freq(train$Raison.Reco.clean[train$Note.Reco == "0"])
#merge les deux tables
freq.all <- merge(word.freq.pos, word.freq.neg, by ='word', all = T )
#clean'up
freq.all$freq.x[is.na(freq.all$freq.x)] <- 0
freq.all$freq.y[is.na(freq.all$freq.y)] <- 0
#calculer la difference
freq.all$diff <- abs(freq.all$freq.x - freq.all$freq.y)
head(freq.all[order(-freq.all$diff), ])
#NDSI
#smoothing term
alpha <- 30
#ndsi
freq.all$ndsi <- abs(freq.all$freq.x - freq.all$freq.y) / (freq.all$freq.x + freq.all$freq.y + 2*alpha)
head(freq.all[order(-freq.all$ndsi), ])
#number of feature in the matrix
num.features <- 500
#sort the values
freq.all <- freq.all[order(-freq.all$ndsi), ]
############################################## tf-idf ###################################################################################
#cast to a string
freq.all$word <- as.character(freq.all$word)
#construct the tf matrix
tf <- t(apply(t(train$Raison.Reco.clean), 2, function(x) str_count(x, freq.all$word[1:num.features])))
idf <- log(length(train.index)/colSums(sign(tf[train.index, ])))
idf[is.infinite(idf)] <- 0
tf.idf <- as.data.frame(t(apply(tf, 1, function(x) x * idf)))
colnames(tf.idf) <- freq.all$word[1:num.features]
##################################################################### Random forest #############################################
# train random forest classifier
ndsi.forest <- randomForest(tf.idf[train.index, ], as.factor(train$Note.Reco[train.index]), ntree = 100)
My question here is : is there anyway to elliminate the stop words or to make less sparse the tf-idf matrix?
Thank you in advance