The evolution of democratic participation has been significantly influenced by digital transformation, especially with the emergence of social media platforms such as X, Facebook, and Instagram. These platforms provide an expansive data source for analyzing public opinion, making traditional reliance on scheduled campaign events less significant. However, translating online sentiments into accurate election predictions remains challenging due to discrepancies between digital discourse and real-world outcomes. Traditionally, voters relied on scheduled campaign events to hear candidates' manifestos Nowadays, candidates and their supporters are turning more to social media to share their manifestos and aspirations, allowing for broader and more immediate communication [1-2]. This shift has turned social media into a robust data source for analyzing public opinion. However, there is often a discrepancy between online sentiment and actual election outcomes, as candidates with low online ratings can still win due to various influencing factors.

To bridge this gap, it is crucial to understand voters' thoughts and emotions accurately. Computational Intelligence (CI) within Artificial Intelligence (AI) focuses on creating adaptive systems that learn from and adjust to complex, dynamic environments. Inspired by human cognitive processes, CI comprises methods like Neural Networks (NN), Fuzzy Logic (FL), Genetic Algorithms (GA), and Machine Learning (ML), deep learning (DL), and swarm intelligence (SI) [3]. These techniques empower systems to process and analyze data, recognize patterns, and make informed decisions amidst uncertainty or incomplete information. CI applications span optimization problems, data mining, expert systems, and autonomous decision-making, highlighting its capacity to evolve and learn from experience [4-5]. Machine Learning (ML) underpins CI by creating intelligent systems that can learn, reason, and solve problems. ML models are trained to generalize patterns from observed data Thomas, Eyoh, by improving performance with more examples and experiences [6-7].

ML processes are classified into supervised learning (trained on labeled data), unsupervised learning (identifying patterns in unlabeled data), and reinforcement learning (learning through interaction and feedback). Hybrid systems integrating ML with other CI techniques, such as fuzzy logic or evolutionary algorithms, yield robust and adaptive intelligent systems. Deep Learning (DL), a subset of ML, utilizes artificial neural networks with multiple layers to learn hierarchical data representations. Key types of deep neural networks include Convolutional Neural Networks (CNNs) for image-related tasks and Recurrent Neural Networks (RNNs) for sequential data [8-9]. DL's scalability and the concept of transfer learning enhance its efficiency and effectiveness in various applications [10]. Despite challenges like the need for substantial labelled data, DL continues to advance, exemplified by transformer-based models [11].

The conception of transformer models by Vaswani, revolutionized Natural Language Processing (NLP) and extended their influence on fields like computer vision and robotics [12-13]. Transformer models, utilizing self-attention mechanisms, excel in capturing contextual relationships within data intelligence [14]. Transformer models have various types tailored to a specific task. The Bidirectional Encoder Representations from Transformers (BERT) by Delvin, and its variant Robustly optimized Bidirectional Encoder Representations from Transformers (RoBERTa) by Liu, approaches have been pivotals in shaping the landscape of natural language processing (NLP) models [15-16]. Others include DistilBERT, a distilled version for faster deployment and the ALBERTa model which introduces parameter reduction techniques for improved efficiency. Variants like BERT and its robust version RoBERTa have set new benchmarks in NLP. The Vision Transformer (ViT) further demonstrated transformers' versatility by challenging CNNs in image-related tasks.

Recent studies demonstrate the growing application of transformer-based models to sentiment analysis and political discourse mining. Talaat, showed that BERT-based and hybrid transformer architectures can achieve strong performance in social-media sentiment classification, highlighting their ability to capture contextual linguistic patterns [17]. In the electoral domain, Pandiangan and Wijayakusuma, applied BERT to sentiment analysis of social-media discourse surrounding the 2024 Indonesian presidential election, demonstrating the relevance of transformer models for analyzing voters’ online opinions [18].

Khatavkar, her extended this line of research through a multilingual transformer-based approach to political tweets, addressing contextual variations and linguistic complexities in political discourse [19]. Similarly, Babu, investigated political sentiment in X (formerly Twitter) comments using transformer models, confirming the effectiveness of contextual language representations for political sentiment classification [20]. More recently, Nguyen and Văn employed XLM-RoBERTa with parameter-efficient fine-tuning for multiclass political sentiment analysis and addressed class imbalance through oversampling and hyperparameter optimization [21].

Collectively, these studies establish the suitability of transformer-based architectures for political sentiment analysis while also revealing persistent challenges related to linguistic ambiguity, multilingual discourse, and imbalanced sentiment classes. RoBERTa, an optimized version of BERT, refines the pre-training process by eliminating the next sentence prediction objective and introducing dynamic masking. This approach enhances RoBERTa's ability to understand a wider range of linguistic patterns and improve generalization. RoBERTa's extensive pre-training on diverse datasets enables it to excel in various NLP tasks, making it a powerful tool for sentiment analysis and election prediction. However, limited attention has been given to integrating these capabilities with election-outcome prediction within the Nigerian context, particularly at the sub-national level.

This study bridges the above-mentioned gap to build on existing transformer-based approaches by applying RoBERTa to X discourse associated with the 2023 Nigerian presidential election in Akwa Ibom State, with the aim of establishing a data-driven relationship between public sentiment and electoral outcomes. Transformer models have redefined Natural Language Processing (NLP) by efficiently capturing contextual relationships within textual data. While existing literature highlights their success in sentiment analysis and classification, this research explores their application in election outcome prediction, emphasizing the unique challenges of sentiment polarity and class imbalance.

This study employs a transformer-based framework using RoBERTa for predicting election outcomes. The RoBERTa model was chosen for its robust pretraining and ability to handle diverse linguistic patterns. RoBERTa was selected because it provides several improvements over earlier transformer-based architectures such as BERT, DistilBERT, and ALBERT. Unlike the original BERT model, RoBERTa removes the next-sentence prediction objective, employs dynamic masking during training, utilizes substantially larger training corpora, and applies optimized pretraining strategies that improve contextual language understanding. These improvements enable RoBERTa to capture subtle semantic relationships, informal expressions, abbreviations, and contextual dependencies commonly found in social media texts. Since election-related tweets often contain sarcasm, slang, abbreviations, and rapidly evolving political narratives, RoBERTa provides a more robust representation of public opinion than conventional transformer models, making it particularly suitable for sentiment-driven election prediction. The proposed methodology workflow for transformer-based election outcome prediction using RoBERTa and electorate opinion is shown in Figure 1.

In this work, a novel approach to predicting election outcomes using electorate opinions was presented to leverage the capabilities of the RoBERTa model for natural language processing. The model was trained on a comprehensive dataset of public sentiment, expressed through various channels such as social media and surveys, to classify opinions as positive or negative towards candidates and political parties. The implementation was carried out using Python, leveraging on PyTorch and Hugging Face Transformer libraries. Google Colab was used for computational efficiency and data visualization. The results of the study revealed key insights into the performance of the RoBERTa model in sentiment-based election prediction. Pie Chart of Election_Tweet_Data sentiment distribution is given in Figure 4. The pie chart in Figure 4 illustrates the sentiment distribution in the dataset, highlighting the proportions of positive (40.8%), negative (38.8%), and neutral (20.4%) sentiments among the processed tweets. Figure 5 illustrates the sentiment distribution, with a balanced representation of positive and negative classes critical for model fairness. Yet, the neutral class remains underexplored, which could provide additional insights into voter indecision or ambivalence.

Table 5: Model Performance across Epochs

The accuracy and loss trends across epochs plot are presented in Figure 7. This illustrates the trends in accuracy, training loss, and validation loss across different epochs, showing stable accuracy with slight fluctuations and consistent low loss values throughout the training process. Table 6 gives RoBERTa model's election prediction analysis across different epochs interval (Observations). The Table summarize the prediction analysis of the RoBERTa model across different epochs. Figure 8 gives the plots of both the training and validation loss, as well as the accuracy, F1 score, precision, and recall across the epochs. The training and validation losses are shown on the left y-axis, while the accuracy, F1 score, precision, and recall are shown on the right y-axis. The sentiment classification metrics result is presented in Table 7 and graphically represented in Figure 9. The bar chart illustrates the sentiment classification metrics for the "Positive" and "Negative" classes. Precision, recall, and F1 score are represented for each class. This visualization highlights the performance differences across the metrics and classes.

Table 7: Sentiment Classification Metrics

Analysis of Epoch 3 on performance metrics indicate that the model exhibits an accuracy of 92.26%, with an F1 score of 0.5519. The precision and recall metrics for this epoch show a moderate balance, with precision at 0.6392 and recall at 0.4856. Classification report indicate that the overall accuracy is high at 92%. For the NEGATIVE class, the precision, recall, and F1-score are robust at 0.95, 0.97, and 0.96 respectively. However, the POSITIVE class struggles, with precision at 0.64, recall at 0.49, and an F1-score of 0.55. This indicates that while the model is proficient in predicting negative sentiments, it has difficulty accurately identifying positive sentiments. Analysis of Epoch 10 on performance metrics indicates that the accuracy slightly drops to 90.41%, and the F1 score for the POSITIVE class remains low at 0.1869. The precision and recall for this class are 0.5584 and 0.1123, respectively, which indicates a significant imbalance. Classification report shows that the NEGATIVE class continues to perform well with an F1-score of 0.95. However, the POSITIVE class still underperforms, with a significant drop in recall to 0.11, leading to an F1-score of 0.19. This highlights a challenge in the model’s ability to generalize well across both sentiment classes, particularly for the POSITIVE class.

Analysis of Epoch 35 and Epoch 100 on the performance metrics shows that the analysis of later epochs (35 and 100) reflects a progressive deterioration in the model's capacity to predict POSITIVE class effectively, indicated by the low F1 scores and a gradual decrease in accuracy, reaching 88.82% by Epoch 100. It is observed that the training loss remains relatively low, but the validation loss fluctuates, suggesting potential overfitting. The model’s decreasing recall and precision for the POSITIVE class signal a need for additional tuning or possibly a different approach to class balancing or model architecture. Specifically, it is observed that the model consistently excelled in predicting negative sentiments, with an F1-score peaking at 0.96. This suggests strong feature extraction for this class, likely due to more distinguishable patterns in negative language. Performance for positive sentiments lagged behind, with an F1-score of 0.55, particularly in later epochs, reflecting the model's struggle to accurately predict positive sentiments. This discrepancy points to the need for targeted interventions such as data augmentation or weighted loss functions to address class imbalance.

Generally, the accuracy decreases slightly over time, with precision and recall for the POSITIVE class showing notable declines as the epochs progress, suggesting potential overfitting or class imbalance issues. Training over multiple epochs revealed diminishing returns, with metrics like precision and recall showing instability beyond Epoch 10. This emphasizes the importance of early stopping and dynamic learning rate adjustments during training Analysis of sentiment classification revealed that the model excels in predicting negative sentiments (F1-score: 0.95), while performance for positive sentiments remains suboptimal. Table 7 and Figure 8 present the performance metrics for the sentiment classification model across two classes: Positive and Negative. The model achieved a higher precision, recall, and F1-score for the Negative class, with values of 0.95, 0.97, and 0.96, respectively, indicating exceptional performance in identifying and classifying negative sentiments. In contrast, the Positive class demonstrated lower metrics, with a precision of 0.64, a recall of 0.49, and an F1-score of 0.55, highlighting a need for improvement in the model's ability to accurately classify positive sentiments. This disparity suggests an imbalance in class performance and emphasizes the importance of addressing it to enhance the overall effectiveness of the model. The incorporation of dropout layers and adaptive learning rates helped mitigate overfitting to some extent but was insufficient to sustain performance across extended training epochs. Hyperparameter tuning and additional regularization techniques remain areas for further exploration. The framework’s ability to translate online sentiment into actionable election predictions underscores its utility for political strategists and analysts. However, the discrepancy between online discourse and actual voter behaviour requires additional contextual analysis, such as geographical or demographic segmentation.

This study highlights the potential of transformer-based architectures like RoBERTa for sentiment-driven election prediction, offering significant contributions to political analysis. The model demonstrated robust performance in predicting negative sentiments but faced challenges with positive sentiment classification, reflecting a need for targeted improvements. Key findings indicate that early learning stages yield high accuracy, but overfitting becomes a concern with extended training. The RoBERTa model demonstrates strong performance in detecting negative sentiments while struggles with positive sentiment prediction across the epochs in this study. Despite high overall accuracy, the disparity between the NEGATIVE and POSITIVE class performances highlights an area for improvement for future works, possibly through advanced techniques like data augmentation, re-sampling, or fine-tuning with more balanced datasets.

The declining F1 scores and recall for the POSITIVE class as training progresses suggest that adjustments in the training process or model hyperparameters could enhance the model’s generalization capabilities, particularly in sentiment analysis tasks. By leveraging the RoBERTa model's capabilities in handling large-scale textual data and capturing nuanced sentiments, this framework aims to provide accurate and timely predictions of election outcomes based on real-time public opinion

The proposed framework underscores the importance of leveraging social media data to bridge the gap between online discourse and real-world outcomes. By refining these methodologies, the research offers a path toward more transparent, accurate, and actionable electoral predictions that can serve as valuable tools for political strategists, media organizations, and policymakers. The performance highlights its potential for providing actionable insights to political stakeholders. However, challenges such as class imbalance and overfitting were noted, particularly in predicting positive sentiments. Future research will integrate text, images, and geospatial data for richer analyses. Oversampling can be used to enhance model generalization, and these will provide deeper insights into voter behaviour.

This study adhered to ethical standards by safeguarding data privacy and anonymity in line with the General Data Protection Regulation and Nigeria’s Data Protection Regulation, ensuring that no personally identifiable information from Twitter data was exposed. Also, data authenticity was ensured by filtering bots, spam, and disinformation, thereby preserving the integrity of the electoral discourse analyzed.

The study was conducted in compliance with institutional research ethics guidelines.

All data were collected from publicly available tweets in line with Twitter’s Developer Policy. No personally identifiable information (PII) was stored; only anonymized tweet IDs and metadata were retained. Data were securely stored on encrypted drives and used solely for academic purposes

This research is supported by TetFund through Institutional Based Research grant 2025 on the project, “High-Precision Sentiment-Based Election Prediction in Emerging Democracies: A Hybrid Intelligence Approach.

The authors declare no conflict of interest. The funders had no role in the study’s design; in the collection, analyses, or interpretation of data; in the writing of the manuscript; or in the decision to publish the results.