aspect sentiment classification

This work addresses the challenge of acquiring high-quality annotated data for fine-grained opinion analysis tasks—such as Aspect Sentiment Triplet Extraction (ASTE) and Aspect-Category Opinion-Sentiment (ACOS)—which are hindered by high annotation costs and substantial human effort, particularly in multi-domain settings. To mitigate these limitations, the authors propose an automated labeling and arbitration framework grounded in large language models (LLMs), integrated with a declarative annotation pipeline. This approach significantly reduces inconsistencies arising from manual prompt engineering while achieving high inter-annotator agreement on both ASTE and ACOS tasks. By minimizing reliance on human annotators and lowering data construction costs, the method enhances the reliability and scalability of cross-model annotations, thereby facilitating broader practical deployment across diverse domains.

This study addresses the scarcity of high-quality annotated data for German aspect-based sentiment analysis (ABSA) and the unclear impact of annotation sources on model performance. It presents the first systematic comparison of annotation quality among experts, students, crowdworkers, and large language models (LLMs) in the German ABSA context. The authors construct a gold-standard dataset through expert re-annotation and evaluate the effectiveness of each annotation type on two core tasks: aspect category sentiment analysis (ACSA) and aspect term and sentiment detection (TASD). Leveraging state-of-the-art models—including BERT, T5, and LLaMA—with both fine-tuning and instruction-based prompting, the experiments demonstrate that expert annotations yield significantly higher consistency and downstream task performance. The study also quantifies the trade-offs of using LLM-generated and non-expert annotations under resource-constrained conditions, highlighting their practical feasibility alongside inherent limitations.

This work addresses the limitations of traditional sentiment analysis, which relies on discrete categorical labels and struggles to model fine-grained, continuous emotional dimensions—such as valence and arousal—and their manifestation at the aspect level. To overcome this, the authors propose a novel approach for dimensional aspect-based sentiment analysis. It first employs a Transformer encoder to jointly predict aspect-specific valence and arousal scores. Second, it innovatively leverages large language models (LLMs) to generate synthetic Russian-language affective descriptions, thereby enhancing multilingual dimensional sentiment regression, particularly for low-resource languages. Finally, the method fine-tunes a decoder-based LLM to enable end-to-end structured extraction of aspects, opinions, categories, and their corresponding sentiment scores. Evaluated on SemEval-2026 Task 3, the approach significantly improves both sentiment score prediction and structured extraction performance for Russian and other low-resource languages.

This work addresses multilingual aspect-based sentiment analysis (ABSA) in实体 retail settings. We introduce and publicly release the first large-scale, manually annotated multilingual customer review dataset for retail—comprising 10,814 samples across eight aspect categories with corresponding sentiment polarities—filling a critical gap in multilingual ABSA benchmark resources for the retail domain. Leveraging this dataset, we systematically evaluate GPT-4 and LLaMA-3 on the joint task of aspect identification and sentiment classification, employing domain-specific prompt engineering and rigorous experimental validation using fine-grained retail corpora. Results show both models achieve accuracy above 85%, with GPT-4 significantly outperforming LLaMA-3 across all key metrics. Our contribution includes: (1) a high-quality, multilingual, retail-specific ABSA benchmark dataset; and (2) an empirically grounded performance baseline for large language models in vertical-domain multilingual ABSA tasks.

Aspect-Sentiment-Opinion Triplet Extraction (ASTE) lacks annotated resources for Slavic languages, particularly Polish, which has no publicly available dataset. Method: We introduce the first Polish ASTE dataset, covering two domains—hotels and e-commerce—and strictly adhering to the standard English ASTE format to ensure cross-lingual comparability. The dataset is manually annotated with fine-grained sentiment structures and released under a CC-BY-NC license. Contribution/Results: Using this resource, we conduct the first systematic evaluation of two mainstream ASTE paradigms and two Polish large language models, revealing critical performance bottlenecks of existing methods on Slavic languages. This work fills a key gap in low-resource, fine-grained sentiment analysis for Slavic languages and establishes a benchmark dataset and empirical foundation for future multilingual ASTE research and model development.

This work addresses the high cost of manual annotation in complex aspect-based sentiment analysis (ABSA) tasks such as aspect sentiment quadruple prediction (ASQP). To mitigate this, the authors propose LA-ABSA, a framework that systematically leverages large language models (LLMs) as annotators, guided by only a few human-provided examples through in-context learning to generate high-quality labeled data for fine-tuning lightweight downstream models. This approach substantially reduces reliance on both extensive human annotation and expensive LLM inference. Evaluated on five benchmark datasets—including SemEval Rest16—the method achieves strong performance, attaining an F1 score of 49.85 on ASQP, which closely approaches the in-context learning performance of Gemma-3-27B (51.10) while significantly lowering computational overhead.

This study addresses the lack of fine-grained, dynamic perspectives in existing research on sentiment analysis of multi-round peer review comments. Focusing on 11,063 accepted papers from *Nature Communications*, the authors constructed a manually annotated corpus of approximately 5,000 sentences and employed deep learning models—including LCF-BERT-CDM—for aspect-level sentiment classification, achieving a Macro-F1 score of 82.65%. Integrating text clustering with statistical analysis, the work systematically uncovers, for the first time, how sentiment evolves across review rounds for key aspects such as “Experiments,” “Research Significance,” and “Results Analysis.” The findings reveal that as the number of review rounds increases, the proportion of positive sentiment rises while negative sentiment declines, and aspect-level sentiment scores exhibit a significant negative correlation with the total number of review rounds.

This study addresses the scarcity of aspect-based sentiment analysis (ABSA) resources for low-resource languages like Czech, particularly the lack of datasets annotated with opinion terms and effective cross-lingual transfer methods. To bridge this gap, the authors construct the first Czech ABSA benchmark dataset in the restaurant domain, featuring opinion-term annotations and supporting three levels of task complexity. They systematically evaluate a range of Transformer and large language models (LLMs) under monolingual, cross-lingual, and multilingual settings, and propose an LLM-driven translation–label alignment strategy to enhance cross-lingual transfer. Experimental results demonstrate that the proposed approach significantly outperforms baseline methods on Czech ABSA, while also revealing limitations of current models in capturing fine-grained opinion expressions, thereby establishing a new benchmark for sentiment analysis in low-resource languages.

Existing aspect sentiment triplet extraction (ASTE) systems lack fine-grained validation mechanisms, often yielding predictions that appear locally plausible but are globally invalid. To address this limitation, this work proposes FiVeD, a novel framework that introduces, for the first time, an adjustable fine-grained verification mechanism. FiVeD jointly optimizes multiple objectives through multi-task learning, including validity classification, quality scoring, error type identification, and diagnostic rationale generation. Leveraging large language models, the framework constructs pseudo-labels enriched with quality scores and justifications, enabling effective filtering or re-ranking of candidate triplets. Evaluated across multiple ASTE baseline models, FiVeD achieves an average F1 improvement of 3.53 points, substantially enhancing both extraction accuracy and system robustness.

This study addresses the challenge of scarce labeled data in sentiment analysis for software engineering, where off-the-shelf sentiment analysis tools often underperform. It presents the first systematic evaluation of various zero-shot learning (ZSL) approaches—including embedding-based methods, natural language inference, TARS, and generative models—on this task, leveraging an expert-defined sentiment label taxonomy. The authors compare these ZSL methods against fine-tuned Transformer models under diverse label settings. Experimental results demonstrate that certain ZSL approaches achieve macro F1 scores comparable to those of fine-tuned models, substantially reducing reliance on annotated data. Error analysis further reveals that subjective labeling practices and confusion between polarity and factual content are primary sources of misclassification.