Meta-XAI: Explainable AI Techniques for Meta Learners with Domain-wise Explanations of Sarcastic Text

Recently, the insights of Explainable Artificial Intelligence (XAI) techniques in automatic sarcasm detection have contributed in understanding the underlying working of the machine learning models, facilitating the identification of the factors influencing the model’s performance while ensuring data privacy and security. Predominantly, meta-models primarily focusing on single domain datasets are used in the literature requiring to deal with challenges of change in data, increasing size of the model, and so on. A baseline Meta-Self-Learner (MSL) provides stable performance while dealing with overfitting and underfitting problems. Authors in this research have used two gold standard datasets, Twitter and MUStARD (MUltimodal SARcasm Dataset) consisting of two different domains, general tweet texts and dialogue speeches respectively, to perform cross-domain analysis of sarcastic texts using MSL and compared the performance with Logistic Regression as meta-model. Further, granular explanations of the model results are provided with Local Interpretable Model-agnostic Explanations (LIME) and Shapley additive explanations (SHAP) techniques at local and global levels to understand the reasons behind the model’s performance and identify the best suitable model across domains. The findings show MSL showed improvement of 8% and 6% of accuracy on cross-domain experiments while ensuring compliance with data privacy and security standards. The explanations with LIME and SHAP have significantly contributed in understanding the best and worst performing models across-domains.