Open research questions in Sentiment Analysis and Opinion Mining

What the literature leaves open

• It remains unclear what specific clusters of AIGC design learning creators are focusing on and what kinds of attitudes, positive or negative, they hold towards these different clusters within the field.

  Creators’ perceptions and attitudes toward using generative artificial intelligence: Exploring posts and comments related to AIGC design learning on a Chinese social media platform with a mixed-method approach · 2026 · DOI

• Future research could focus on building a decision‑making framework that balances three dimensions — accuracy, cost, and explainability. Future research could explore the application of LLMs' fine-grained sentiment extraction and reasoning capabilities to the early identification of mental health risks in social media texts, and construct a specialized sentiment analysis evaluation system for mental health by integrating psychological frameworks of emotion classification.

  The Technological Evolution of Sentiment Analysis—A ComparativeStudy from SVM to Large Language Models · 2026 · DOI

• Sentiment analysis with frozen pre-trained language model (PLM) backbones has become a common paradigm, yet the practical benefit of explicit domain adaptation remains unclear, particularly when backbones encode varying degrees of target-domain knowledge.

  Is Domain Adaptation Always Helpful? A Frozen-Backbone Study of Cross-Domain Sentiment Transfer · 2026

• Future research should focus on validating the proposed framework using larger real- world datasets, exploring LLM-based ABSA approaches for improved contextual understanding, and conducting cross-country studies to assess the generalizability of findings across different tourism contexts.

  Hybrid ABSA–C5.0 framework for interpretable classification of tourist perceptions in digital destination services · 2026 · DOI

• This interpretability offers a unique to ABSA research and helps explain contribution satisfaction dynamics in under-studied Turkish second- hand marketplaces.

  Understanding Aspect-Sentiment Drivers of Overall Ratings in Second Hand Marketplace Apps through Text Analytics and Regression Analysis · 2026 · DOI

• Future work may explore hierarchical classification architectures, tar- geted data augmentation strategies, contrastive pre-training approaches, and cross-platform validation on other Indonesian super-app ecosystems such as Grab, Tokopedia, and Shopee to assess the generalizability of the proposed framework. Third, the current study assumes that conjunction-aware segmentation produces semantically valid sentence bounda- ries; however, segmentation quality was not evaluated independently and therefore warrants further investigation. By decomposing multi-opinion reviews into independently classifiable sentence units, the proposed framework addresses a key limitation of conventional document-level sentiment analysis, namely the loss of opinion granularity in reviews containing multiple sentiment tar- gets.

  Sentence-Level Sentiment Analysis of Indonesian App Reviews Using IndoBERTweet · 2026 · DOI

• As illustrated in case (b), when the text contains positively biased expres- sions such as “important steps” while the visual modality fails to offer effective complementary information, cross- modal interactions are insufficient to rectify the sentiment bias introduced by the text, resulting in prediction errors.

  Cmcl-kmse: adaptive multimodal aspect-based sentiment analysis leveraging cross-modal multi-anchor contrastive learning and knowledge-guided multi-view semantic enhancement · 2026 · DOI

• Nazir MK, Faisal CN, Habib MA, Ahmad H (2025) Leveraging multilingual transformer for multiclass sentiment analysis in code-mixed data of low-resource languages.

  Cross-lingual sentiment analysis via multimodal transformer fusion and lightweight deep ensemble learning framework · 2026 · DOI

• Educational aspect-based sentiment analysis (ABSA) can support course improvement, but public aspect-labeled student feedback remains scarce because educational reviews are private, institution-specific, and expensive to annotate.

  A Controlled Synthetic Benchmark for Educational Aspect-Based Sentiment Analysis · 2026

• This study proposes a new approach to e-commerce product combining sentiment analysis with a LSTMmodel based on DL. The system is structured to efficiently handle large volumes of user-generated Amazon review data, using systematic preprocessing. The LSTM architecture which is comprised of embedding, stacked LSTM, and dense layers is trained on the Binary Cross-Entropy loss and the AdamOptimizer to ensure effective convergence. Experimental results show that performance is excellent, with an acc of 98.42, and F1score of 98.70, which indicate a high predictive performance and balanced classification performance. This is further indicated by the ROC-AUC score that indicates better class separability. Evaluations against both conventional ML and competing DL models demonstrate the superiority of the suggested method for extracting complicated sequential patterns from textual input. This approach is great for e-commerce systems because it improves suggestion quality by considering user sentiment, which improves customization and decision making. The proposed framework may be used in real-time recommendation systems and the creation of intelligent e-commerce apps since it is generally scalable, precise, and efficient. The proposed model has some disadvantages even when it is doing well. It may not pick up on indicators of user activity because of its dependence on textual review data. Also, TF-IDF is ineffective in terms of the ability to extract in-depth semantic meaning. The model has high computing resource demands, which might compromise scalability in real-time. In future studies, greater contextual understanding can be realized through a combination of transformer-based models such as BERT and multi-modal information such as user behavior and product metadata. The computational cost and real-time efficiency of large-scale e-commerce platforms may be further optimized. REFERENCES 1. K. Dixit, “Predictive Analytics in Business Intelligence for Sales Forecasting,” Int. J. Adv. Res. Sci. Commun. Technol., vol. 60, no. 3, p. 981, Sep. 2023, doi: 10.48175/IJARSCT-12750G. in 2024 2. M. S. Rahman, T. D. Sarkar, U. T. Mitasha, M. S. Mia, and S.

  An Efficient Deep Learning Framework for Real-Time Product Recommendation in E-Commerce · 2026 · DOI

• User-item interaction data E-commerce image dataset AFETLER (Adaptive Fusion Model) Multi-source recommendation data High accuracy (95.8%) with reduced bias and variance Captures linear & nonlinear relations effectively (AUC: 0.94) High accuracy (95%+) and good recall in image-based…

  An Efficient Deep Learning Framework for Real-Time Product Recommendation in E-Commerce · 2026 · DOI

• approach that improves personalization by leveraging user review polarity. • Enhances recommendation accuracy by capturing contextual meaning and sequential patterns in textual feedback using deep learning. • Improves learning from imbalanced e-commerce datasets, leading to more reliable and unbiased predictions. • Achieves consistently performance compared to existing ML and DL baselines across all evaluation metrics. superior • Strengthens real-time recommendation quality by linking user sentiment with product directly relevance. • Demonstrates high scalability and robustness for deployment in large-scale e-commerce environments. For e-commerce real-time product recommendation, the proposed strategy is innovative since it combines sentiment analysis with an LSTM-based DL model. Rather of relying on generic recommendations, it improves customization by using the emotion of user reviews. With the help of SMOTE, the model is able to handle class imbalance and accurately capture sequential textual patterns, resulting in more accurate predictions. Its better performance over conventional ML and DL approaches across all of the evaluation measures justifies its explanation. Its ability to provide scalable, accurate, and sentiment-aware product suggestions has been proven to greatly improve customer satisfaction. A. Structure of Paper The rest of the paper is organized as follows: Section II reviews the relevant literature. A thorough description of the recommended method is given in Section III. The experiments and their findings are presented in Section IV. Lastly, Section V concludes and outlines future directions. II. LITERATURE REVIEW The following sections include machine learning, product recommendation systems and a literature review on techniques and algorithms used to develop better recommendation systems. images. To handle noisy and Siddharth and Sariki, (2025) present a multimodal deep learning approach, in this study that uses a fusion-based model to merge structured attribute information with product incomplete metadata, we apply preprocessing steps such as one-hot encoding and class balancing, while CNNs are used to extract RichVisual features. Our experiments show that the fusion model consistently outperforms image-only and attribute-only baselines, reaching 84.2% ± 2.1 test accuracy and a macro F1score of 0.87 across five folds. Goranthala et al., (2025) use of the GPU-YOLO Ensembled Classifier is essential in reducing the usual problems of bias and variance that older classifiers have. The Am. J. Interdiscip. Innov. Res.

  An Efficient Deep Learning Framework for Real-Time Product Recommendation in E-Commerce · 2026 · DOI

• algorithms that frequently encounter data sparsity and cold-start issues. The contextual significance of user evaluations is also not captured by them. Machine learning techniques make text more adaptable, but it's impossible for fully comprehend complex semantic patterns sequential presently. When modeling dependencies in evaluation data, deep learning, and LSTM networks in particular, perform better. Additionally, user evaluations include valuable sentiment data that is sometimes overlooked. As a result, there's an opportunity to enhance suggestion quality by merging deep learning with sentiment analysis. The need for more precise, personalized, individualized, and context-aware recommendation algorithms for modern e-commerce platforms have inspired the research. This paper's primary contributions are as follows: similarity are conventional • Introduces a sentiment-driven recommendation approach that improves personalization by leveraging user review polarity. • Enhances recommendation accuracy by capturing contextual meaning and sequential patterns in textual feedback using deep learning. • Improves learning from imbalanced e-commerce datasets, leading to more reliable and unbiased predictions. • Achieves consistently performance compared to existing ML and DL baselines across all evaluation metrics. superior • Strengthens real-time recommendation quality by linking user sentiment with product directly relevance. • Demonstrates high scalability and robustness for deployment in large-scale e-commerce environments. For e-commerce real-time product recommendation, the proposed strategy is innovative since it combines sentiment analysis with an LSTM-based DL model. Rather of relying on generic recommendations, it improves customization by using the emotion of user reviews. With the help of SMOTE, the model is able to handle class imbalance and accurately capture sequential textual patterns, resulting in more accurate predictions. Its better performance over conventional ML and DL approaches across all of the evaluation measures justifies its explanation. Its ability to provide scalable, accurate, and sentiment-aware product suggestions has been proven to greatly improve customer satisfaction. A. Structure of Paper The rest of the paper is organized as follows: Section II reviews the relevant literature. A thorough description of the recommended method is given in Section III. The experiments and their findings are presented in Section IV. Lastly, Section V concludes and outlines future directions. II. LITERATURE REVIEW The following sections include machine learning, product recommendation systems and a literature review on techniques and algorithms used to develop better recommendation systems.

  An Efficient Deep Learning Framework for Real-Time Product Recommendation in E-Commerce · 2026 · DOI

• One key area of future work is the long-term impact of in- creasing user agency. Does using Alexandria yield any long- term attitudinal or behavioral changes? The changes exe- cuted by an extension such as Alexandria have the potential to directly impact the visibility of polarizing content. A lon- gitudinal evaluation would also allow us to understand what values people configure over time. Do people revert to en- gagement ranking, confirming platforms’ argument that en- gagement is indeed preferred by their users? Does a value- based ranking reduce engagement? Are users happy with these trade-offs? Our approach currently only operates on the web client for social media platforms. Mobile applications of centralized social media platforms are generally restricted from client modification of this sort. This restriction limits our ability to extend our extension to the mobile experience. Other more open platforms, such as Mastodon and Bluesky, are much more obvious candidates if mobile interventions are desired since feed ranking can be performed with server-side cus- tomization. Future work should investigate how Alexandria customization can be extended to mobile devices. Finally, future work must investigate the governance structures to support a marketplace of value and what moderation ap- proaches can be implemented.

  Alexandria: A Library of Pluralistic Values for Realtime Re-Ranking of Social Media Feeds · 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.

  Multi-model Fusion for Emotion Detection in Text: A Stacking and Majority Voting Approach · 2026 · DOI

• outline clear directions for future research, including multilingual multi-label classification, and the integration of more efficient XAI methods such as SHAP. extension, dataset V.

  Explainable Multi-Emotion Mental Health Classification from Twitter Emotions Dataset Using Bidirectional GRU with LIME · 2026 · DOI

• Although SGMCL demonstrates robust performance, two primary limitations suggest directions for future research. First, the model exhibits sensitivity to the quality of linguistic inputs. The system relies on external priors (parsers, SenticNet) and pre-trained encoders. Consequently, performance may degrade in two scenarios: (1) Complex Structures, where standard parsers fail to capture longrange dependencies in nested clauses; and (2) Informal Contexts, where slang terms (e.g., “meh”) lack coverage in static lexicons or robust embeddings in BERT. While our Multi-Channel Contrastive Learning (MCL) module effectively aligns features for standard expressions, its efficacy is naturally bounded by the semantic quality of these inputs. Future work will explore latent graph learning and domainadaptive pre-training to reduce reliance on static priors and bridge semantic gaps. Second, while our span-based greedy inference effectively acts as a structural regularizer to ensure high precision and structural coherence, it may be restrictive in specific complex scenarios. Specifically, by enforcing strict boundary alignment between sentiment, aspect, and opinion terms, the current strategy prioritizes structural integrity over exhaustive recall. This design choice, while beneficial for reducing noise in high-confidence extraction, inherently limits the model’s ability to recover certain heavily overlapping or loosely connected triplets. Future work could explore constrained global inference mechanisms that can better balance this precision-recall trade-off without compromising the structural coherence established by the encoder. Addressing these limitations will not only enhance the robustness and scope of ASTE models but also contribute to bridging the gap between structured prior knowledge and more autonomous, data-driven representation learning.

  Span labeling with sentiment-aware GCN and multi-channel contrastive learning for aspect sentiment triplet extraction · 2026 · DOI

• Therefore, the extent to which the effectiveness of WM-SAR is model-agnostic remains to be verified, and future work should conduct comparative experiments with a more diverse set of LLMs. 2 Limitations This study has three main limitations. First, the LLMs used in this work are mainly limited to the GPT-4.

  World model inspired sarcasm reasoning with large language model agents · 2026 · DOI

• This research presents an effective and robust approach for sentiment analysis using a Bidirectional Long Short-Term Memory (BiLSTM) model. The primary objective of the study was to classify customer reviews into positive and negative sentiments by leveraging deep learning techniques and a well-structured Natural Language Processing pipeline. The proposed methodology integrates key steps such as data cleaning, text preprocessing, tokenization, sequence padding, and word embedding, followed by model construction using the BiLSTM algorithm. This systematic approach ensures that raw textual data is transformed into meaningful representations, enabling accurate sentiment classification. 1713 World Journal of Advanced Research and Reviews, 2026, 30(01), 1703-1716 The experimental results demonstrate that the proposed model achieves exceptionally high performance, with accuracy reaching nearly 100% on both training and validation datasets. The model also maintains a strong balance between precision and recall, resulting in a high F1-score, which confirms its effectiveness even in the presence of class imbalance. The training and validation graphs further indicate that the model converges quickly, with minimal loss and stable learning behaviour. The confusion matrix analysis shows that almost all predictions are correct, with negligible misclassification, highlighting the reliability of the system. A key strength of this research lies in the use of the BiLSTM algorithm, which processes textual data in both forward and backward directions. This bidirectional learning mechanism allows the model to capture contextual relationships between words more effectively than traditional machine learning models. As a result, the proposed system significantly outperforms conventional approaches such as Naive Bayes and Support Vector Machines, which lack the ability to understand sequential dependencies in text. Additionally, the use of regularization techniques such as dropout and early stopping plays a crucial role in improving model generalization and preventing overfitting. The incorporation of class weighting further ensures balanced learning across different sentiment classes. Overall, the developed system proves to be scalable, efficient, and suitable for real- world applications such as customer feedback analysis, product review monitoring, and opinion mining. 5.1. Future Work Despite achieving high accuracy and strong performance, there are several opportunities for future enhancement. The current model focuses on binary classification; however, it can be extended to multi-class sentiment analysis to capture more detailed sentiment categories such as neutral or highly polarized sentiments. Future work may also include aspect-based sentiment analysis, which can identify sentiments related to specific product features or attributes, providing deeper insights for business decision-making. Furthermore, integrating advanced transformer-based models such as BERT can further improve contextual understanding and overall model performance. In addition, the development of real-time sentiment analysis systems capable of processing streaming data from social media platforms can enhance the practical applicability of the model. Extending the framework to support multilingual sentiment analysis is another promising direction, enabling the system to handle diverse datasets across different languages. These enhancements will further strengthen the scalability, adaptability, and real-world usability of the proposed sentiment analysis system.

  Deep learning-based sentiment analysis of customer reviews using bidirectional LSTM · 2026 · DOI

• Although the proposed system achieves good performance, there are several areas for further improvement. Future research can focus on implementing advanced deep learning models such as BERT (Bidirectional Encoder Representations from Transformers) to achieve better contextual understanding and higher accuracy. Handling multilingual and code-mixed data is another important area, as customer reviews may contain multiple languages. Improving sarcasm detection is also necessary, as existing models may fail to correctly interpret sarcastic or ambiguous statements. In addition, implementing real-time sentiment analysis can enhance the system by enabling immediate processing of customer feedback. Furthermore, incorporating Explainable AI techniques such as LIME and SHAP can improve model transparency by explaining how predictions are made. Overall, future enhancements will focus on improving model accuracy, handling complex data, and making the system more efficient and reliable for real-world applications.

  Sentiment Analysis of Uber Customer Reviews Using Machine Learning and Deep Learning Techniques · 2026 · DOI

• To enhance operational feasibility and pedagogical value, future work will explore lightweight model variants through techniques such as compression and knowledge distillation, while also integrating explainable AI (XAI) modules—e. Future work will focus on incorporating multimodal emotion recognition, expanding dataset diversity, and optimizing model efficiency for large-scale, real-time educational applications.

  A study on learners’ emotion classification based on an improved convolutional neural network algorithm in online teaching and learning · 2026 · DOI

• systems), but results are broadly applicable to any scenario involving textual opinion mining on social media. applications (e.g., II. LITERATURE REVIEW Early work by Pang and Lee demonstrated that machine learning classifiers (SVM, Naïve Bayes) could achieve high accuracy (≈83%) on movie review sentiment data using unigram features. Subsequently, many researchers have explored varied approaches: lexicon-based, ML-based, and deep learning. A. ML and Preprocessing Symeonidis et al. conducted a comprehensive study on Twitter sentiment preprocessing. They evaluated multiple features (n-grams, TF-IDF) and classifiers (Linear SVM, Naïve Bayes, CNN). Their results showed that a CNN using word embeddings outperformed traditional ML, achieving higher classification accuracy. Similarly, Huq et al. applied k-NN and SVM on they reported Twitter data with n-gram features; the moderate accuracy (58–80%) and highlighted © 2026 The Author(s). Published by IJCOPE Journal. Website: https://ijcope.org/ 2 International Journal of Creative and Open Research in Engineering and Management finding importance of feature selection. Amolik et al. analyzed movie-related tweets with feature vector approaches, that SVM yielded better recall/sensitivity than Naïve Bayes. In contrast, Liao et al. compared a simple CNN (with word2vec) against SVM on Twitter, concluding that CNN achieved higher accuracy in Twitter sentiment classification. These studies illustrate that deep models often surpass shallow classifiers when sufficient data is available. B. Deep Learning Methods Convolutional and recurrent neural networks have become popular. For instance, a CNN+word2vec model on a Twitter dataset achieved balanced precision/recall of 88.7%. Another study by Zheng et al. used a hybrid bidirectional RNN on mixed datasets (Sogou news, Yelp, Douban reviews), achieving accuracy up to ~97% on some data. Zhao et al. proposed a weaklysupervised deep embedding model for Amazon product reviews, reaching 87.9% accuracy. These advances show the power of DL architectures for sentiment analysis. However, DL performance can depend heavily on data size and representation quality. C. Contextual Embeddings (BERT and Transformers) found Recent approaches use large pre-trained language models. Basarslan and Kayaalp compared word embedding methods (Word2Vec, GloVe, BERT) and classifiers on multiple review datasets (IMDb, Yelp, Twitter). They that models using BERT embeddings "have the best performance" over TF-IDF or static embeddings. For example, BERT-based models achieved up to 94–98% accuracy on benchmarks, outperforming traditional ML by 5–10 percentage points. This agrees with the broader literature: contextual models capture nuances of language that simple bag-ofwords methods miss [6]. D. Research Gaps is a Despite numerous studies, gaps remain. Many papers evaluate one or two datasets in isolation, without crosslack of systematic domain analysis. There comparison of modern Transformer-based models versus classic methods on review data. Furthermore, few studies examine hybrid pipelines that combine multiple feature types or adapt pretrained models specifically for social reviews. Our work addresses these gaps by benchmarking diverse approaches on the same datasets and proposing an integrated method. social media ISSN: 3108-1754 (Online) Volume 02 Issue 04 April-2026 | Impact Factor: 3.5 III. METHODOLOGY This section details the proposed sentiment analysis framework. We adopt a hybrid pipeline combining advanced text representation with a neural classifier. Key components are: (i) text preprocessing, (ii) feature extraction, (iii) classification model, and (iv) training loss. The overall system architecture is illustrated conceptually in Fig. 1. [← Fig. 1: Proposed BERT+BiLSTM Framework →] Fig. 1. Proposed sentiment analysis framework: input text → BERT encoder → BiLSTM → Softmax classifier. A. Preprocessing Raw text reviews are first cleaned by lowercasing, removing URLs, user mentions, and non-alphanumeric characters. Standard NLP preprocessing such as and tokenization, stemming/lemmatization are applied to normalize input. This step reduces noise in social media text, consistent with prior studies. stop-word removal, B.

  Sentiment Analysis of Social Media Reviews: A Machine Learning and Deep Learning Approach · 2026 · DOI

• This study successfully developed and validated a sentiment analysis model that effectively captures the emotional tone of King Abdullah II's speech at the United Nations. The results revealed a significant distribution of sentiments, with the Call to Action category comprising 16% of the speech, underscoring a strong emphasis on mobilizing international support for humanitarian issues. The model achieved an impressive accuracy of 85%, alongside a precision of 83% and recall of 81%, demonstrating its robustness in effectively classifying sentiments. These findings highlight the model's capability to discern nuanced emotional expressions, such as the 30% Empathetic sentiment that fosters compassion and solidarity among audiences, and the 15% Critique of Power, which underscores the need for accountability in global responses. The study’s understanding underline the critical role of emotional appeals in political communication, emphasizing how effectively conveyed sentiments can influence public perception and engagement. However, there are several avenues for future research that could enhance this work. Firstly, expanding the dataset to include a broader corpus of speeches from various leaders would improve the generalizability of the findings and allow for comparative analyses across different political contexts. Additionally, enhancing feature engineering by exploring contextual embedding and sentiment lexicons could further refine model performance, especially in capturing complex sentiments that may be present in political discourse. Moreover, investigating alternative machine learning algorithms and ensemble methods could identify more effective approaches for sentiment classification, potentially increasing accuracy and reducing misclassifications. Conducting longitudinal studies would provide valuable patterns into how sentiments in political speeches evolve over time and their impact on public perception and policy responses. Finally, integrating multimodal analysis, including video and audio data, would enable a comprehensive examination of how non-verbal cues complement verbal sentiments, enriching the understanding of political communication. By addressing these areas, future research can significantly advance the understanding of sentiment in political discourse and its implications for humanitarian advocacy and public engagement, ultimately contributing to more effective communication strategies in addressing global challenges. The proposed framework for conducting nuanced sentiment analysis in political discourse is structured into seven distinct phases, each designed to systematically enhance the understanding of emotional tones within speeches, specifically illustrated through King Abdullah II's address at the United Nations. This comprehensive framework serves as a model for future research in sentiment analysis, emphasizing the importance of methodical data handling and analytical rigor.

  A New Framework for Nuanced Sentiment Analysis in Political Communication: A Machine Learning Approach to the Case Study of His Majesty King Abdullah II's UN Speech · 2026 · DOI

• Cross-sector contradictions—headlines that are positive for one Bangla financial sector but negative for another—confuse the attention mechanism in SSABE-TSCM. The paper identifies this failure mode but does not propose sector-specific attention routing or weighted ensemble methods to handle sector-dependent sentiment polarity reversal.

  SSABE-TSCM: drift-aware and interpretable financial sentiment analysis for low-resource Bangla via adaptive semi-supervised and temporal contrastive modeling · 2026 · DOI

• Crisis-era terminology (e.g., 'digital taka' in early 2023) occurs too seldom in the training set to be learned effectively. Expanding vocabulary coverage via targeted back-translation or synthetic data generation specifically for rare Bangla financial crisis terms is required to maintain macro-F1 above 0.74 during rapidly evolving economic events.

  SSABE-TSCM: drift-aware and interpretable financial sentiment analysis for low-resource Bangla via adaptive semi-supervised and temporal contrastive modeling · 2026 · DOI

Most-cited papers in Sentiment Analysis and Opinion Mining

• Measuring news sentiment · Journal of Econometrics · 2020 · 315 citations
• Deep learning for detecting financial statement fraud · Decision Support Systems · 2020 · 225 citations
• A multimodal approach to cross-lingual sentiment analysis with ensemble of transformer and LLM · Scientific Reports · 2024 · 167 citations
• Sentiment analysis of Chinese documents: From sentence to document level · Journal of the American Society for Information Science and Technology · 2009 · 162 citations
• Sentiment analysis methods, applications, and challenges: A systematic literature review · Journal of King Saud University - Computer and Information Sciences · 2024 · 159 citations
• Sentiment Analysis in the Age of Generative AI · Customer Needs and Solutions · 2024 · 155 citations
• Multimodal Sarcasm Detection via Hybrid Classifier with Optimistic Logic · Journal of Telecommunications and Information Technology · 2022 · 119 citations
• Performance Metrics for Multilabel Emotion Classification: Comparing Micro, Macro, and Weighted F1-Scores · Applied Sciences · 2024 · 111 citations
• Sarcasm detection using machine learning algorithms in Twitter: A systematic review · International Journal of Market Research · 2020 · 111 citations
• Sentiment analysis of public services for smart society: Literature review and future research directions · Government Information Quarterly · 2022 · 104 citations

Most recent work

• Enhancing Sentiment Analysis and Error Prediction in Lawyer Questioning Through Legal NLP Models: A Mixed-Methods Study in Australian Courtrooms · Journal of Mixed Methods Research · 2026
• SSABE-TSCM: drift-aware and interpretable financial sentiment analysis for low-resource Bangla via adaptive semi-supervised and temporal contrastive modeling · Frontiers in Artificial Intelligence · 2026
• ViSP: A PPO-enhanced framework for multimodal sarcasm generation with contrastive learning · Neurocomputing · 2026
• Cross-lingual sentiment analysis via multimodal transformer fusion and lightweight deep ensemble learning framework · Knowledge and Information Systems · 2026
• Alexandria: A Library of Pluralistic Values for Realtime Re-Ranking of Social Media Feeds · Proceedings of the International AAAI Conference on Web and Social Media · 2026
• The Application of BERT in Sentiment Analysis of IMDB Movie Reviews · Journal Of Social Research · 2026
• A Survey on AI-Powered Meme Generator for Company Advertisements · International Journal of Scientific Research in Engineering and Management · 2026
• Text Optimized XLNet Based Sentimental Analysis for Opinion Mining Towards Products From Twitter Data · International Journal of Computational Intelligence Systems · 2026
• Enhancing fine-grained sentiment detection in Chinese text through context-aware attention mechanism and emotional vocabulary fusion · Progress in Artificial Intelligence · 2026
• Beyond the stars: predicting guest satisfaction through NLP and machine learning of hotel reviews · Journal of Hospitality and Tourism Technology · 2026

Related topics in Computer Science