Optimizing Semantic Deep Forest for tweet topic classification

Nowadays, topic detection from Twitter attracts the attention of several researchers around the world. Different topic classification approaches have been proposed as a result of these research efforts. However, four of the major challenges faced in this context are the use of handcrafted features, the use of Deep Learning algorithms with so many parameters, the fact that their performance is still limited and the lack of sufficient labeled datasets. We propose, Semantic Deep Forest (SDF), a topic classification approach that incorporates contextual Word2vec, WordNet and Deep Forest to detect topic from Twitter accurately. Moreover, extensive parameter sensitivity analysis were conducted to fine-tune the parameters of SDF for our Tweet topic classification task to achieve the best performance. We conducted experiments on three benchmark datasets with standard evaluation scenarios. Experimental results show that: (1) the proposed contextual word2vec models can be successfully used for tweet topic classification and outperform existing state-of-the-art embedding model; (2) The proposed SDF improve the accuracy of tweet topic classification and outperform existing state-of-the-art classification approaches; (3) the proposed SDF does not require huge amount of labeled data in order to achieve good performance, which is the lack in the majority of the state-of-the-art approaches.