(The pdf of code isn't that helpful. A small reproducible example would go a long way to help us help you)
How many features are generated from the one column of text?
What is the memory utilization when not run in parallel? Are you exhausting memory since the use of parallel processing has multiplied the total memory needs by 3?
I'd suggest moving away from the formula method for train (or anything else that uses random forests). If the predictors are categorical (and we can't tell from your example code in pdf; it's hard to read), it probably takes longer to run.
Use method = "ranger" instead of the randomForest package and turn off ranger's internal parallelism.
Finally, be aware that the call to train is fitting 91 random forest models. It's not really "a" random forest model.