Lie detection using RoBERTa with bidirectional gated recurrent unit (BiGRU) optimized by enhanced running game optimization (ERGO)

Identifying deceptive statements accurately remains challenging due to subtle linguistic cues, contextual ambiguity, and domain variability. The suggested model includes several state-of-the-art elements, such as RoBERTa-based contextual word embeddings, a bidirectional gated recurrent unit (BiGRU) for sequential modeling, and an enhanced running game optimization (ERGO) algorithm for hyperparameter optimization. RoBERTa offers highly contextualized textual input representations, whereas the BiGRU layer focuses on these representations by learning bidirectional temporal dependencies. ERGO optimizer also improves the performance of the models by dynamic optimization of important parameters that include learning rate, dropout rate and batch size, and size of hidden units, which enhances convergence stability. The model was tested on two benchmark datasets, namely, the LIAR dataset, containing 12 836 fact-checked political statements, and the generalized deception dataset, which contains 125 590 deceptive and truthful texts in a wide variety of fields, including reviews, phishing emails, job scams, political statements, and news articles. A tenfold cross-validation was used to ensure sound and unbiased analysis. The results of the proposed ERGO-RoBERTa-BiGRU model have been shown to be very impressive with an accuracy of 98.87%, a precision of 98.75%, a recall of 98.38%, and an F1-score of 98.56%, which are much higher than all base and comparative models. Ablation studies also helped establish the fact that every architectural element contributes to the general performance.