Detection of causality in sentences: a comparative study of ensemble and non-ensemble approaches across embedding paradigms with explainable AI

Detection of causality in sentences aims to identify causal relationships expressed in natural language text. This task is critical for applications in finance, healthcare, safety, business risk management, scientific literature, and policy making. In this study, we present a comprehensive comparative analysis of causality detection using ensemble and non-ensemble approaches. The ensemble framework integrates Random Forest, Naive Bayes, XGBoost, and Linear Support Vector Machines with soft and weighted-soft voting. To evaluate effectiveness, experiments are grouped into three categories: (i) traditional embeddings (TF-IDF, Bag-of-Words, Word2Vec, FastText, GloVe), (ii) deep learning architectures (CNN, LSTM, GRU, BiLSTM, CNN-GRU, CNN-LSTM), and (iii) transformer-based embeddings (BERT, RoBERTa, BART, MiniLM, DistilBERT, DeBERTa). For broader comparison, non-ensemble baselines include Support Vector Machines, Random Forest, XGBoost, and Neural Tangent Kernel for traditional embeddings, while SVM-RBF is used for deep learning and transformer embeddings. Evaluations are conducted on four publicly available datasets using F1-score, precision, accuracy, recall, and ROC-AUC with tuning hyperparameter (for non-ensemble approach) and 10-fold cross-validation. Results show that ensemble-based approaches consistently outperform non-ensemble methods, achieving F1-scores of 99.98%, 98.38%, 99.72%, and 99.99% on the four datasets. Finally, explainable AI techniques (LIME, Anchor, Counterfactuals) are employed to interpret model decisions, providing insights into causality detection mechanisms.