Align and Conquer: An Ensemble Approach to Classify Aggressive Texts from Social Media

The phenomenal proliferation of social media platforms has facilitated the spontaneous sharing of expressions, opinions, and emotions in public spaces. Unfortunately, with this rapid rise, these mediums have been repeatedly used to spread propaganda, excite religious and political violence, jeopardize social harmony and disseminate other aggressive content. The pernicious societal effects of such undesired content have become a severe concern for tech giants and government bodies. Studies revealed that the majority of such activities are performed via texts written in regional languages. Developing an automatic system to identify and classify aggressive texts in resource constraint languages like Bengali is monumental. This work proposes an ensemble-based classifier model to compensate for this deficiency. A corpus of 10095 annotated texts (5095 for non-aggressive and 5000 for aggressive) is developed to train the system. Moreover, aggressive texts are classified into four fine-grained categories: political, gendered, verbal and religious using hierarchical annotation schema. This work also investigates 21 standard classifier models developed based on Convolutional Neural Network (CNN), Bidirectional Long Short Term Memory (BiLSTM) and Gated Recurrent Unit (GRU) with different embedding models (i.e., Word2Vec, GloVe, FastText) and ensemble strategies. All the models are trained, tuned and tested on the developed dataset (EATxtC-Extended Aggressive Text Corpus). The experimental result exhibits that the ensemble of CNN and GRU (i.e. CNN+GRU) outperformed the other baseline models with acquiring the highest weighted f1-score of 89.55% (coarse-grained) and 83.77% (fine-grained).