Word Level Approach for Tweets Classification based on its Content

Twitter has become the largest microblogging platform where users can interact between each other expressing opinions, thoughts and feelings related to any topic or source of news in a compressed 280 character message, called tweet. Hashtags are popular keywords used to label these tweets according to its content. This work tries to nd out if the usage of hashtags to label tweets with similar content is accurate enough. To do so, tweets from di erent popular hashtags have been retrieved and processed in order to have a dataset with a content as close to reality as possible. Several embedding methods and learning algorithms have been studied to classify tweets from di erent hashtags based on the content. Results showed that the best performance is achieved when using the Tf-idf embedding method and support vectors machine. The learning algorithm obtained a precision around 90% for classi cation on 10 classes and above 70% when dealing with 100 classes trained on datasets of only 13680 and 143067 samples respectively. The results also indicated that BoW and Tf-idf methods outperformed other state of the art methods for other natural language processing tasks, such as GloVe or Word2Vec. Outgoing