An Auto-ML Framework for social media domain feature construction to detect Sarcasm

Recently, Automated Machine Learning (AutoML) has shown considerable growth and industry application in the domain of Machine Learning. AutoML main purpose is to improve the learning task by saving time and effort for tasks like preprocessing, feature engineering, model selection, hyperparameters, and model architecture. Generally, AutoML lacks the ability to produce the generalized features for specific or general tasks like sentiment and sarcasm classification, therefore, to overcome the shortcoming of general feature. We proposed generalized feature extraction algorithms Implicit Incongruity (IIA) and Explicit Incongruity (EIA), the main aim is to observe the newly extracted incongruity feature and integrated into AutoML DeepConcat model at search pipeline. Additionally, proposed model will be evaluate using the preprocessing plan with various levels, where each level represents a single preprocessing task like cleaning, the model performance varies with from level to level. BiLSTM is the best model, which is the outcome of the framework, it was selected among five deep learning models during search pipeline. The AutoML DeepConcat framework automates the model selection by concatenating these features at hidden layers of the models. Further, the performance was optimized by evaluating the models with hyperparameters like dropout and learning rate. The proposed framework all-features method outperformed other features methods like pragmatic and incongruity.