TL;DR
A hybrid model, termed SE_LLM_ST, is introduced, which leverages recent advancements in large language models (LLMs) and transformer-based architectures to effectively capture both contextual and sequential information vital for Persian emotion recognition.
Understanding emotional tone and sentiment in text has driven significant advancements in Natural Language Processing (NLP), particularly in sentiment analysis and emotion detection. This study addresses the challenge of developing effective NLP tools for low-resource languages, focusing on the Hausa language. By leveraging ensemble methods and pre-trained transformer models like BERT and XLM-R, along with traditional classifiers such as Logistic Regression, SVM, Naive Bayes, Random Forest, and XGBoost, we aim to improve sentiment analysis and emotion detection for Hausa text. Utilizing a balanced sentiment dataset (9,958 samples) and a complex multi-label emotion dataset (19,757 samples across 11 categories), we benchmark individual classifiers, voting ensembles, and deep contextual models. For sentiment analysis, a Hard Voting Ensemble of TF-IDF-vectorized base learners achieved a highly competitive F1-score of 0.8748. However, Transformer models significantly outperformed traditional baselines, with Multilingual BERT (mBERT) achieving a peak F1-score of 0.8983. In the multi-label emotion detection task, individual traditional models struggled with label sparsity, yielding low Subset Accuracy scores (2.88% to 8.30%) and moderate Micro-F1 scores. Standard Hard Voting ensembles further underperformed due to discrete prediction conflicts. To resolve this, a Probability-based Majority Voting mechanism with calibrated thresholding (0.3) was introduced, boosting the Micro-F1 to 0.3825 and reducing the Hamming Loss to 0.1967. Ultimately, XLM-RoBERTa emerged as the superior architecture, achieving a Subset Accuracy of 0.1545, a Micro-F1 of 0.4275, and the lowest Hamming Loss of 0.1804. This research establishes a rigorous benchmark for Hausa NLP, highlighting the indispensable role of subword tokenization, contextual embeddings, and threshold calibration in handling the morphological richness and multi-label complexities of low-resource languages
: Emotion recognition plays a crucial role in enabling intelligent human–computer interaction, yet research in low-resource languages such as Urdu remains limited, particularly in multimodal settings. This study proposes a multimodal deep learning framework for Urdu emotion recognition by integrating speech and text modalities. The approach leverages transformer-based models, namely wav2vec 2.0 for audio representation and MuRIL for text representation, combined using a late fusion strategy for classification. Experiments were conducted on the UMED dataset, consisting of 8269 multimodal instances across five emotion classes. The proposed multimodal model achieved an accuracy of 0.701 and an F1-score of 0.6915, outperforming unimodal baselines, where the audio-only and text-only models achieved accuracies of 0.6681 and 0.5085, respectively. Furthermore, the proposed approach surpasses the existing UMEDNet benchmark, demonstrating the effectiveness of transformer-based feature extraction and multimodal late fusion for Urdu emotion recognition. The results highlight the complementary nature of speech and text modalities and demonstrate that independently learned modality-specific classifiers combined through decision-level fusion can improve emotion recognition performance in low-resource languages. However, the performance improvement over alternative fusion strategies was relatively modest, indicating that more advanced multimodal interaction mechanisms may further enhance recognition performance.
Emotion detection from social media is crucial for understanding human emotions across languages. However, for low-resourced languages such as Amharic, the lack of annotated data makes this task challenging. Additionally, most current models use black-box methods that obscure whether predictions rely on linguistically meaningful cues. To address these gaps, this study proposes a multi-label emotion classification model for Amharic by fine-tuning AfroXLMR. To enhance transparency, we integrate explainable artificial intelligence (XAI) into the framework. We compiled and annotated a new dataset of 22,000 unique social media comments across eight emotion categories for training, validation, and testing. The data was split into 80% for training, 10% for validation, and 10% for testing. The proposed model achieved a recall of 87% and a Hamming loss of 0.08. To interpret its predictions, we applied Local Interpretable Model-agnostic Explanations (LIME). We also evaluated the model against several state-of-the-art baselines, including XLM-R base, mBART, BiLSTM, LSTM, CNN, and AfriBERTa. The results show that our approach outperformed each baseline, achieving F1-score improvements of 5% over XLM-R base, 3% over mBART, 5% over BiLSTM, 7% over LSTM, 9% over CNN, and 2% over AfriBERTa. Bootstrapped statistical significance testing confirms that these improvements are robust and not attributable to random variation. In conclusion, the fine-tuned AfroXLMR model demonstrates promising performance in Amharic multi-label emotion classification. Building on this success, next steps could involve exploring more advanced fine-tuning strategies and expanding our datasets to strengthen both performance and the model's ability to generalize across diverse Amharic contexts.
Emotion detection from textual data is a key challenge in natural language processing (NLP), playing an important role in applications such as sentiment analysis, human-computer interaction, and psychological evaluation. Growing utilization of social networks and online portals leads to the creation of a huge amount of reviews and ratings. Analyzing users’ and customers’ reviews and opinions are so important for governments and businesses. While recent advances have primarily leveraged transformer-based architectures for this task, we propose a novel approach by employing Deep Operator Networks (DeepONets), originally designed for learning operators in scientific computing, to model the mapping between textual representations and emotional states. In this study, we extract high-dimensional semantic embeddings from text using a pre-trained sentence transformer model and feed these embeddings into a DeepONet architecture for emotion classification. The primary contribution of this work lies in architectural innovation. To the best of our knowledge, this is the first application of DeepONet to textual emotion recognition, introducing a fundamentally different perspective on function approximation in language understanding tasks. Results obtained from experiments on benchmark emotion-labeled datasets indicate that our proposed model attains performance and results comparable to related baselines, with notable generalization capabilities across emotion categories. To ensure a fair and comprehensive evaluation, we assessed the performance of the proposed model using widely adopted classification metrics, including accuracy, precision, recall, and F1-score. We used two datasets and the results were around 80 percent on one dataset for all named metrics, and around 88 percent on the other. The findings suggest that DeepONet can serve as a robust alternative framework for capturing and modeling complex relationships in natural language processing tasks and opens new avenues for operator-based learning in text analysis.
It is critical for emotion-aware multichannel adaptation techniques to have high accuracy in terms of recognizing emotions, particularly in English-speaking learners' cases. Recurrent neural network-based methods, aggregation-based approaches, and conventional multimodal fusion techniques are known to have some drawbacks related to time dependency, dynamicity of emotional transitions, and inter-speaker variance robustness. In an attempt to address these limitations, we propose a task-specific multimodal framework that employs an emotion-structured transformer encoder coupled with a progressive emotion distillation strategy. What should be noted first about this paper is that its contribution lies in neither technique employed, since they were previously utilized in different works. The key idea behind the proposed approach is the combination of existing emotion recognition techniques into a multimodal pipeline that facilitates improved emotion recognition and robust representation learning. Estimation of the speaker-independent representation, emotion-aware representation, and improved emotion recognition are achieved through employing emotion-prior masked attention, emotion-gated feature transformation, multi-head attention pooling, and boundary-aware auxiliary supervision in ESTE. Further improvement in training the recognition model is obtained by incorporating emotionally ambiguous samples progressively and a distillation process based on soft labels, in addition to auxiliary tasks such as valence-arousal regression and boundary detection. Improved results are empirically observed when compared with selected baselines using emotion recognition benchmark datasets. As a conclusion, an interpretative teaching strategy generation method relying on the rules of teaching guidelines is suggested.
Emotion classification on social media provides valuable insights into public sentiment, but the performance of existing models is often limited by corpus size and linguistic variability. This research presents a transfer learning approach to analyze the benchmark EmoT corpus of 4401 emotion-labeled Indonesian Tweets, combined with a task-specific data augmentation strategy to enhance model generalization. Statistical analysis is performed using a linear mixed model of 10-fold cross validation folds, with folds modeled as a random intercept to control for within-fold variation. Results reveal that both Model and Augmentation Strategy have a significant effect on Accuracy, Macro F1, and Weighted F1 metrics. The top-performing model-strategy combination is IndoRoBERTa with augmentation via one-phase back translation, achieving a Weighted F1 score of approximately 0.859. These results highlight the effectiveness of integrating transfer learning with textual data augmentation for emotion classification in low-resource languages and suggest promising directions for future research in natural language processing.