Latest Models in Sentiment Analysis: A Short Literature Review

Introduction Sentiment analysis—the computational study of opinions, emotions, and attitudes in text—has shifted dramatically over the past three years. Where the field once leaned on lexicons and classical machine learning, recent work is now dominated by transformer architectures, large language models (LLMs), and multimodal systems. This short review summarises the main directions in research published between 2024 and 2025. Transformer-Based and Hybrid Deep Learning Models Transformer encoders such as BERT and RoBERTa remain the strongest baselines for sentence- and document-level sentiment classification, but recent work has focused on hybridising them with recurrent and convolutional components to capture both contextual and sequential signals. Madhurika and Naga Malleswari (2025) proposed SentiNet, a hyperparameter-optimised deep architecture for customer-review classification that outperformed standalone CNN and LSTM baselines. Similarly, Atlas et al. (2025) combined BiGRU and LSTM networks with rich linguistic preprocessing to improve product-review polarity detection. A systematic review by Mohammadi et al. (2025) covering 2019–2024 confirmed that hybrid models—particularly CNN-BiLSTM and attention-augmented RNNs—have become the dominant paradigm for social-media sentiment tasks. Large Language Models for Sentiment Analysis The rise of generative LLMs has reshaped expectations of what sentiment models can do. Kirtac and Germano (2024) compared OPT (a GPT-3-class model), BERT, FinBERT, and the Loughran-McDonald dictionary across nearly one million financial news articles, reporting that the OPT-based system reached 74.4% directional accuracy in predicting stock returns and clearly outperformed the smaller encoders. In the tourism domain, Vatsa et al. (2024) showed that fine-tuned GPT-4o and GPT-4o-mini surpassed BERT on hotel-review star-rating prediction, indicating that few-shot adaptation of frontier LLMs can beat domain-trained encoders. Rusnachenko et al. (2024) extended this line further by applying chain-of-thought prompting with Flan-T5 to targeted sentiment analysis in Russian news, achieving state-of-the-art results and demonstrating that explicit reasoning steps add measurable value over zero-shot prompting. Aspect-Based Sentiment Analysis (ABSA) Aspect-based sentiment analysis, which targets opinions toward specific entities or features, has benefited from both transformer fine-tuning and explainability research. Nazir and Rashid (2024) developed an explainable ABSA framework that combined transformer encoders with attribution methods to make aspect-level predictions interpretable for end users. On the architectural side, recent ABSA work increasingly relies on graph convolutional networks layered over transformer embeddings to model syntactic dependencies between aspect terms and opinion words (Tayal et al., 2024). Multimodal Sentiment Analysis A major frontier is multimodal aspect-based sentiment analysis (MABSA), where text is fused with images, audio, or video. Sun and Zhu (2024) introduced the Multilayer Interactive Attention Bottleneck Transformer (MIABT), which uses a dynamic gating mechanism to align image and text features while limiting cross-modal information flow for efficiency. Zou et al. (2025) proposed the Target-oriented Cross-Modal Transformer (TCMT), integrating textual, visual, and OCR-derived features through graph- and CNN-enhanced transformer blocks. Cross-lingual multimodal ensembles have also emerged: Hossain et al. (2024) combined neural machine translation with Twitter-RoBERTa-based sentiment models across Arabic, Chinese, French, and Italian, reporting that ensemble fusion meaningfully reduces translation-induced sentiment drift. Discussion and Research Gaps Three trends stand out across these studies. First, frontier LLMs are now competitive with or superior to fine-tuned encoders, but only when properly adapted through prompting or parameter-efficient fine-tuning. Second, hybrid and graph-augmented transformers remain attractive when computational budgets are constrained or when aspect-level structure matters. Third, multimodality and cross-linguality are the most active research frontiers, although both still struggle with modality alignment, noisy visual cues, and translation artefacts. Persistent gaps include sarcasm and irony detection, low-resource language coverage, and the lack of standardised benchmarks for explainable and multimodal ABSA. Conclusion Sentiment analysis in 2024–2025 is defined by the convergence of LLM-driven reasoning, hybrid transformer–recurrent architectures, and multimodal fusion. Continued progress will likely depend less on raw model scale and more on alignment, interpretability, and robust cross-domain generalisation.

References

References