This study investigates the application of advanced fine-tuned Large Language Models (LLMs) for Turkish Sentiment Analysis (SA), focusing on e-commerce product reviews. Our research utilizes four open-source Turkish SA datasets: Turkish Sentiment Analysis version 1 (TRSAv1), Vitamins and Supplements Customer Review (VSCR), Turkish Sentiment Analysis Dataset (TSAD), and TR Customer Review (TRCR). While these datasets were initially labeled based on star ratings, we implemented a comprehensive relabeling process using state-of-the-art LLMs to enhance data quality. To ensure reliable annotations, we first conducted a comparative analysis of different LLMs using the Cohen’s Kappa agreement metric, which led to the selection of ChatGPT-4o-mini as the best-performing model for dataset annotation. Our methodology then focuses on evaluating the SA capabilities of leading instruction-tuned LLMs through a comparative analysis of zero-shot models and Low-Rank Adaptation (LoRA) fine-tuned LlaMA-3.2-1B-IT and Gemma- 2-2B-IT models. Evaluations were conducted on both in-domain and out-domain test sets derived from the original star-ratings-based labels and the newly generated GPT labels. The results demonstrate that our finetuned models outperformed leading commercial LLMs by 6% in both in-domain and out-domain evaluations. Notably, models fine-tuned on GPT-generated labels achieved superior performance, with in-domain and outdomain F1-scores reaching 0.912 and 0.9184, respectively. These findings underscore the transformative potential of combining LLM relabeling with LoRA fine-tuning for optimizing SA, demonstrating robust performance across diverse datasets and domains.

This study investigates the performance of traditional Machine Learning (ML) models for Sentiment Analysis (SA) in the Turkish language, addressing the growing need for robust Natural Language Processing (NLP) techniques in low-resource languages. We compare the efficacy of Logistic Regression (LR), Support Vector Machine (SVM), Random Forest (RF), and Naive Bayes (NB) classifiers on two recently developed large-scale Turkish sentiment datasets: “Vitamins and Supplements Customer Review” (VSCR) and “Turkish Sentiment Analysis version I” (TRSAv1). Our research leverages advanced hyperparameter tuning techniques to optimize these models, providing benchmarks and practical insights for Turkish SA. The study contributes to the field by establishing performance baselines on new open-source datasets, enhancing understanding of NLP models for agglutinative languages, and analyzing factors affecting classifier performance in binary and multi-class scenarios. Our findings reveal the strengths and limitations of each model within the context of Turkish language SA, offering valuable guidance for future research and applications in this domain.

With the increasing prevalence of e-commerce plat-forms, understanding customer sentiments expressed in product reviews is crucial for assessing platform and product performance. However, traditional sentiment analysis (SA) models often struggle with low-resource languages such as Turkish, particularly due to the language's agglutinative nature and morphological complexity. To address this, several public datasets are now available online. However, these datasets are often annotated using rule-based methods, resulting in poor accuracy for Machine Learning (ML) models. This study addresses these challenges by employing two large Turkish datasets, TRSAv1 and VSCR, and improving their sentiment annotations through advanced Large Language Models (LLMs) such as ChatGPT4o-mini. The reannotation significantly enhanced the performance of traditional ML models-Logistic Regression, Naive Bayes, Random Forest, and Support Vector Machine-especially in the neutral and negative sentiment classes, as evidenced by increased precision, recall, and F1-scores. The findings underscore the importance of thorough data preparation and demonstrate that utilizing LLMs for sentiment annotation in Turkish can lead to more reliable SA, providing a stronger foundation for developing advanced SA tailored to low-resource languages. This study is the first to employ recent LLM for relabelling Turkish customer reviews, highlighting the potential for improved accuracy in SA for low-resource languages.

Behavioural heterogeneity matters in group decision making, but most large language model (LLM) agent deliberation studies do not manipulate it explicitly. We present a reproducible framework for group deliberation grounded in opinion-dynamics theory, in which agents vary in stubbornness, susceptibility, and confidence bound while solving a binary evidence-aggregation task with objective ground truth. The same pipeline supports both a simulation-based pilot study and local language modelbased runs, allowing matched simulation-to-model transfer. In a 2400-run simulation-based pilot, trait composition primarily affects coordination: stubborn-minority mixtures reach operational consensus more often and with lower polarisation, whereas homogeneous anchored groups converge slowest and remain most polarised. We then run two matched 2400-run local language model-based replications with Qwen3.5-35B-A3B and Ministral-3B. Both preserve a separation between agreement and correctness: Qwen reaches consensus 0.829 with final accuracy 0.668, Ministral reaches consensus 0.772 with final accuracy 0.671, and both remain below private-majority accuracy 0.795 and DeGroot accuracy 0.813. Qwen also preserves a clearer trait-conditioned coordination pattern, whereas Ministral weakens it. Overall, the results show that trait-controlled prompting can reshape group coordination, but in the present language model-based setting, it improves coordination more reliably than correctness.

KIDO pairs children’s prompted happy and sad drawings with written reflections, creating a multimodal setting in which the text and image channels provide very different evidence for the same labels. We reanalyze KIDO by comparing Turkish, original English, and LLM-generated Turkish-grounded English text; replicating image-model configurations; and measuring text–image score discrepancy on 1630 drawings from 815 children. Turkish-grounded English is numerically higher than original English across emotion, gender, and educational-level classification under the controlled English protocol. For emotion, the respective three-seed macro-F1 scores are $0.9808\pm 0.0004$ and $0.9798\pm 0.0034$; however, the paired ensemble difference is small, and its 95% child-cluster bootstrap interval includes zero. A blinded, tail-enriched bilingual audit of 120 child-distinct pairs provides source-grounded support for the generated text: 88.3% of ratings receive adequacy scores of 4–5, while major meaning errors occur in 7.5% of ratings and are confined to the deliberately sampled low-similarity tail. Text performance also depends strongly on explicit emotion cues: such cues occur in approximately two-thirds of test reflections, and masking them at evaluation reduces macro-F1 by 0.054–0.073, while performance remains above 0.90. Image-emotion models achieve a substantially lower mean macro-F1, ranging from 0.7683 to 0.7867 across replicated DenseNet121 and ConvNeXt configurations. Cross-modal score differences remain substantial: the Multimodal Discrepancy Index, equivalent to total variation in this binary task, ranges from 0.2300 to 0.2753 across image ensembles. These results show that text preparation, lexical cues, and image-model configuration materially affect conclusions drawn from KIDO and support reporting channel-specific evidence before interpreting multimodal agreement or fusion.