A Cascaded Ensemble Framework Using BERT and Graph Features for Emotion Detection From English Poetry

Researchers have developed complex deep-learning models to extract emotions from poetry, opting for these over lightweight models. However, this approach requires a high volume of resources, which can be a significant limitation. Moreover, they often suffer from overfitting, making them less effective in real-time scenarios. This work introduces a novel cascaded ensemble framework that combines the strengths of BERT and Graph features (CP Net). The framework is designed in a tiered approach, classifying the majority of emotions in poetry, thereby reducing the need to invoke more complex models later on. This strategic arrangement enables efficient resource allocation and minimizes the usage of complex models. The basic model may not classify all so that a later model will classify residual unclassified emotions in the poetry. While the basic models in the first phase classify the majority of emotions in poetry, the remaining unclassified emotions are then passed on to the subsequent stages of the model, where more complex models can further process and classify them. This work performs feature fusion only in case of failed input samples. The performance of the proposed model is evaluated and compared against four baseline models using word embedding models, including Glove and Fast Text. The proposed CP Net model demonstrated exceptional performance, surpassing all other models by achieving an impressive 95% accuracy on the CAPEMO dataset and an outstanding 98% accuracy on the BAPEMO dataset. CP Net achieved state-of-the-art results with a minimum computational time of 0.274 to 0.332 milliseconds, outperforming other models.