Multi-Class and Automated Tweet Categorization

Abstract: Twitter is among the most prevalent social media platform being used by millions of people all over the world. It is used to express ideas and opinions about political, social, business, sports, health, religion, and various other categories. The study reported here aims to detect the tweet category from its text. It becomes quite challenging when text consists of 140 characters only, with full of noise. The tweet is categorized under 12 specified categories using Text Mining or NLP, and Machine Learning (ML) techniques. It is observed that a huge number of trending topics are provided by Twitter but it is really challenging to find out that what these trending topics are all about. Therefore, it is extremely crucial to automatically categorize the tweets into general categories for plenty of information extraction tasks. A large dataset is constructed by combining two different nature of datasets having varying levels of category identification complexities. It is annotated by experts under proper guidelines for increased quality and high agreement values. It makes the proposed model quite robust. Various types of ML algorithms were used to train and evaluate the proposed model. These models have explored over three datasets separately. It is explored that the nature of the dataset is highly non-linear therefore complex or non-linear models perform better. The best ensemble model named, Gradient Boosting achieved an AUC score of 85\%. That is much better than the other related studies conducted.