First, the effectiveness of the models in classifying emotions from FER2013, RAB-DB, and CK+ datasets was evaluated by a 5-fold cross-validation method, which showed that the accuracy varied widely am

The gap

Consensus across the literature

Supporting evidence — 3 representative gaps

• NLP Framework to Safeguard Youngsters Online Using Advanced Transformer-Based Models (2026) · doi

  Even though the current models achieved promising results, it has been recognized that they might not be the absolute best fit for text sentiment analysis and emotion detection. Our work was constrained by several practical limitations: difficulties managing the sheer volume of data, performance dips when classifying a wide spectrum of emotions, the need for substantial time investments, and insufficient computational resources to fully explore and fine-tune state-of-the-art architectures, such as HuBERT, LSTM, and LLMs. In addition to the primary limitations discussed, our team managed several smaller, practical issues. Finding the optimal model architecture for our specific application was a complex selection process that required considerable time. The study faced the persistent difficulty of locating an ethically and technically suitable offensive dataset, a resource crucial for the research’s scope. On the logistical side, typical research bottlenecks emerged, such as minor time losses when multiple experiments had to queue for limited computational power and the demanding effort required for team-based debugging when combining codes from different contributors.

  Keywords: time several practical limitations computational team required even though current models achieved promising recognized absolute
• Multi-model Fusion for Emotion Detection in Text: A Stacking and Majority Voting Approach (2026) · doi

  This study, which is based on Kaggle’s Emotion Text Dataset, has several inherent limitations that should be addressed. Limitations include: • While commonly utilized, the Kaggle Emotion Text Dataset may not capture the complete range of emotions expressed in real-world language. The dataset might be skewed toward specific types of emotional expressions (e.g., joy, rage), resulting in model performance biases. Furthermore, the dataset’s language may not accurately reflect the broad range of speech and writing styles seen in various locations, cultures, and age groups. Emotion detection methods, especially when combined through stacking or voting, may struggle to capture rich emotional context. Sarcasm, irony, and cultural allusions are all subtle clues that might impact emotional reactions, which the models may not completely grasp. As a result, the accuracy of emotional predictions may be compromised in complicated conversational or highly contextual settings. • • • Although stacking models is a strong strategy for enhancing accuracy, it also increases the risk of overfitting, particularly if the individual models in the stack are too complicated or highly connected. This may reduce the total ensemble’s robustness when applied to out-of-sample data or real-world events other than the training set. The majority vote method, while useful for merging predictions, may not necessarily give the best outcomes. If separate models have considerable conflicts, the majority vote may result in inaccurate forecasts. This is especially troublesome when the ensemble has numerous models that are equally confident yet erroneous. • While preprocessing techniques like tokenization, lemmatization, and stop-word removal might help models perform better, they can also eliminate or distort essential emotional cues. Negations (e.g., "not happy") and intensifiers (e.g., "very sad") may lose significance along these processes, resulting in inaccurate emotion identification. The multi-model fusion strategy, particularly stacking, necessitates significant computer resources for both training and inference. For big datasets or real-time applications, this may result in slower processing times, increased memory utilization, and higher operating expenses, making the approach less suitable for deployment in resource-constrained contexts. • • While the models were trained on the Kaggle Emotion Text Dataset, their performance in other emotion-labeled datasets or real-world applications may differ. Different datasets may have distinct emotional distributions, language use, or domain-specific features, making it difficult to generalize the findings without additional validation using varied data sources. • Although the models are designed to categorize emotions from text, they may not always "understand" the underlying emotional context. Emotion identification in text remains mostly focused on surface-level patterns such as keywords and sentence structure, with no deeper psychological or contextual study. As a result, the models may overlook subtler or more complicated emotional states that are not expressly mentioned in the text. Volume 18 (2026), Issue 3 – Page 202 Multi-model Fusion for Emotion Detection in Text: A Stacking and Majority Voting Approach

  Keywords: models emotion emotional text dataset real stacking result kaggle world language model complicated majority datasets
• Facial Sentiment Analysis Using Convolutional Neural Network and Fuzzy Systems (2024) · doi

  First, the effectiveness of the models in classifying emotions from FER2013, RAB-DB, and CK+ datasets was evaluated by a 5-fold cross-validation method, which showed that the accuracy varied widely among different emotion classes and was affected by overfitting.

  Keywords: first effectiveness models classifying emotions datasets evaluated fold cross validation accuracy varied widely among different