The advent of spam on social media platforms has lead to a number of problems not only for social media users but also for researchers mining social media data. While there has been substantial research on automated methods of spam detection on Twitter, research on the lexical content of spam on the platform is limited. A dataset of 301 million generic tweets was filtered through a URL blacklisting service to obtain 7207 tweets containing links to malicious web-pages. These tweets, considered spam, were combined with a random sample of non-spam tweets to obtain an overall dataset of 14,414 tweets. A total of 12 numerical tweet features were used to train and test a Random Forest algorithm with an overall classification accuracy of over 90%. In addition to the numerical features, the text of each tweet was processed to create four frequency-mapped corpora pertaining uniquely to spam and non-spam data. The corpora of words, emoji, numbers, and stop-words for spam and non-spam were plotted against each other to visualize differences in usage between the two groups. A clear distinction between words, and emoji used in spam, and non-spam tweets was observed.
