Search Results (26)

Large language models (LLMs) are increasingly deployed in multilingual settings, yet their safety behavior under Turkish harmful prompts and prompt injection attempts remains insufficiently characterized. This study evaluates the adversarial robustness of 55 open- and closed-source LLMs under paired Turkish and English harmful prompt conditions. We constructed a benchmark of 790 Turkish adversarial prompts, translated the prompts into English for cross-lingual comparison, and applied both prompt sets to the model pool. Model responses were labeled as harmful, harmless, or hallucinatory, and safety was analyzed using safety scores, Turkish–English ranking differences, and inter-rater reliability based on Fleiss’ kappa. The results reveal substantial variation across models. Closed-source systems generally achieved higher safety scores and stronger filtering behavior, whereas open-source and Turkish-oriented models showed a wider performance distribution. GPT-5.4 ranked first in the Turkish tests with a 99.37% safety score but decreased to 96.71% in the English tests, while Qwen3.5:27B ranked first in English with 97.47%. These differences suggest that safety mechanisms are not fully language-invariant. Hallucination also emerged as a distinct safety risk, particularly in Turkish evaluations. The findings indicate that Turkish LLM safety cannot be inferred from general model capability alone and should be assessed through language-specific, culturally aware, and continuously updated adversarial benchmarks. Full article

Background: Psychological distress, particularly anxiety and depression, is highly prevalent among cancer patients, and is associated with impaired quality of life, reduced treatment adherence, and increased mortality risk. Standardized screening instruments, such as the Hospital Anxiety and Depression Scale (HADS), are effective, but face implementation barriers in busy oncology outpatient settings. This cross-sectional study investigated whether BERT-based Natural Language Processing (NLP) analysis of brief patient-generated free texts would correlate with HADS scores in a consecutive cohort of cancer outpatients. Material and Methods: A total of 165 consecutive adult cancer outpatients were enrolled at a tertiary oncology center in Turkey. All participants completed the HADS questionnaire and were asked to write freely about their current emotional state in Turkish. Patient-generated texts were analyzed using a pre-trained Turkish BERT model to derive a continuous BERT Sentiment Score (BSS) and a categorical BERT Sentiment Cluster (BSC) via unsupervised hierarchical clustering. Univariate and multivariate linear regression analyses were performed to examine associations between clinical, demographic, and NLP-derived variables and the logarithmically transformed HADS score. Results: The mean total HADS score was 10.46 (range, 0–33), consistent with a moderate level of psychological distress. In multivariate analysis, two variables were independently associated with HADS scores: female sex (β = 0.20, t = 2.14, p = 0.034), associated with higher HADS scores, and BERT Sentiment Score (BSS) (β = −0.18, t = −2.43, p = 0.016), with higher values corresponding to lower HADS scores. Hierarchical clustering identified two distinct thematic groups: ‘Coping and Fighting Spirit’ (74%), and ‘Hope and Negative Feelings’ (26%); however, cluster membership (BSC) was not independently associated with HADS scores (β = −0.02, p = 0.789). Clinical variables, including cancer stage, diagnosis type, treatment status, and time since diagnosis, also were not independently associated with HADS scores. Conclusions: BERT-based sentiment analysis of brief patient-generated free texts yielded a continuous measure that independently correlated with HADS scores in cancer outpatients, alongside female sex. These findings provide proof-of-concept evidence that NLP-derived sentiment scoring may offer a practical, scalable, and complementary approach to standardized psychological screening in routine oncology care. Full article

This study investigates a target-based sentiment analysis approach on Turkish texts and examines how lexicon-based methods vary depending on language compatibility and translation strategies. The main objective is to accurately identify target-oriented expressions and to compare the performance of different sentiment lexicons within this context. For this purpose, Turkish user reviews obtained from the Turkish school review and evaluation platform were analysed using three lexicon configurations: SentiWordNet applied in its original English form with target-related term translation (SentiWordNet-EN), its fully Turkish-translated version (SentiWordNet-TR), and a native Turkish resource (SentiTurkNet). SentiTurkNet achieved the highest weighted average F1-score of 0.887 (positive-class F1: 0.926; negative-class F1: 0.760), followed by SentiWordNet-EN with a weighted average F1-score of 0.856 (positive-class F1: 0.898; negative-class F1: 0.720), and SentiWordNet-TR with a weighted average F1-score of 0.824 (positive-class F1: 0.868; negative-class F1: 0.679). One of the most significant findings is that using SentiWordNet in its original English form yields better results than the fully translated version, suggesting that the translation process leads to sentiment loss due to the incomplete preservation of sentiment intensity and contextual meaning. These findings carry important implications for sentiment analysis in low-resource languages: where comprehensive native lexicons are unavailable, translating only target-related terms into a language with richer sentiment resources can be more effective than directly translating the entire lexicon. Therefore, it is concluded that in target-based sentiment analysis, not only language compatibility but also the chosen translation strategy plays a critical role. Full article

The rapid growth of user-generated reviews on e-commerce platforms has created a significant decision-making challenge for both consumers and sellers, particularly in morphologically rich low-resource languages such as Turkish. This study proposes a unified artificial intelligence framework for Turkish e-commerce review intelligence by integrating transformer-based sentiment classification, instruction-tuned large language model summarization, and explainable fuzzy logic-based product evaluation within a single end-to-end architecture. A balanced dataset containing 183,333 Turkish reviews was constructed from Trendyol, Amazon Turkey, and Hepsiburada using LLM-assisted annotation and stratified downsampling. Experimental evaluations demonstrated that the fine-tuned BERTurk 128k model achieved a macro F1-score of 0.9243 on the held-out test set. To overcome the limitations of multilingual news-oriented summarization models on informal review text, the framework employed the Turkish instruction-tuned Kumru-2B model together with structured prompt engineering to generate sentiment-aware abstractive summaries. In addition, a Mamdani-type fuzzy inference system was designed to combine sentiment distribution, seller reliability, star ratings, and review volume into an interpretable product-level score. The complete pipeline was integrated into a FastAPI and React-based web platform capable of processing approximately 850 reviews in under 60 s. The findings demonstrate that domain-specific Turkish language models combined with explainable reasoning mechanisms can provide accurate, scalable, and human-interpretable decision support for large-scale e-commerce environments. Full article

Breast imaging reports are commonly recorded as unstructured free-text documents, which limits their secondary use for large-scale clinical analysis, structured information extraction, and clinical decision support. These challenges are particularly important in morphologically rich and low-resource clinical languages, where linguistic variability, inconsistent terminology, and limited annotated corpora may reduce the direct applicability of existing natural language processing (NLP) approaches. This study presents a scoping review of NLP research on breast imaging reports, conducted and reported in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses Extension for Scoping Reviews (PRISMA-ScR). Attention is given to BI-RADS (Breast Imaging Reporting and Data System)-related analytical tasks, methodological trends, dataset characteristics, evaluation practices, and implications for low-resource clinical NLP settings, including Turkish. A comprehensive literature search was conducted across Web of Science, Scopus, PubMed, IEEE Xplore, and Google Scholar in February 2026. To reduce the risk of missing recent studies using transformer- and large language model (LLM)-related terminology, a targeted supplementary search was also conducted. Following screening and eligibility assessment, 39 studies were included in the final synthesis. The findings show that the literature is concentrated mainly on BI-RADS classification/annotation and information extraction tasks. Task-wise, BI-RADS classification/annotation was the most frequent category, followed by information extraction. Methodologically, the reviewed literature shows a shift from rule-based and traditional machine learning approaches toward transformer- and LLM-based methods. LLM-based studies were frequently represented, particularly among recent studies and those identified through the targeted supplementary search; therefore, their observed prominence should be interpreted cautiously. Despite these advances, the literature remains linguistically imbalanced and methodologically heterogeneous. English was the most frequently represented report or dataset language, whereas Turkish breast imaging NLP studies remained limited. Major challenges identified across the reviewed studies include dataset heterogeneity, inconsistent annotation practices, variable evaluation metrics, limited external validation, and incomplete reporting of reproducibility-related details. This scoping review provides a structured synthesis of methodological trends, task categories, dataset characteristics, evaluation practices, and reproducibility-related limitations in breast imaging NLP. Overall, the findings highlight the need for better documented datasets, standardised evaluation practices, transparent reporting, clinically grounded validation, and stronger research efforts for low-resource clinical language settings. Full article

This study presents a controlled empirical and comparative analysis of existing data augmentation techniques for text generation in Turkish, a morphologically rich, low-resource language. A collection of 265 Turkish reading passages for Grades 4 and 5 was augmented using four techniques: paraphrasing with GPT-3.5-turbo (Generative Pre-trained Transformer 3.5 Turbo), back translation (Turkish–English–Turkish and Turkish–French–Turkish) via Google Translate, synonym replacement via GPT-3.5-turbo, and random insertion via GPT-3.5-turbo. Human evaluators assessed the fluency, coherence, grammaticality, logical flow, and naturalness of the augmented datasets. Each augmented dataset, along with the original, was then used to fine-tune a Turkish GPT-2-medium model, which was evaluated using automatic metrics such as BLEU (Bilingual Evaluation Understudy), ROUGE (Recall-Oriented Understudy for Gisting Evaluation), METEOR (Metric for Evaluation of Translation with Explicit ORdering), chrF (CHaRacter-level F-score), BERTScore (Bidirectional Encoder Representations from Transformers Score), and cosine similarity. According to the human evaluation of the original and augmented datasets, the original texts received the highest ratings, followed by those generated through random insertion, paraphrasing, synonym replacement, and back translation variants, with cosine similarity results between original and augmented texts showing a comparable trend; however, the differences between methods were generally small. The results from text generation indicate that models trained on the original dataset generally achieved slightly higher performance across evaluation metrics compared to those trained on augmented datasets. Among the augmented methods, synonym replacement showed marginally better performance, followed by back translation, random insertion, and paraphrasing; however, the differences between methods were small and not statistically significant. Full article

Background/Objectives: Natural language processing (NLP) has emerged as a promising approach for extracting clinically meaningful information from unstructured radiology reports. While most artificial intelligence applications in musculoskeletal imaging focus on image-based analysis, the potential of NLP for urgency assessment in superficial soft tissue ultrasound reports remains underexplored. This study aimed to develop and evaluate an NLP-based triage model to classify superficial soft tissue ultrasound reports according to clinical urgency in orthopedic practice. Methods: A curated dataset of superficial soft tissue ultrasound reports requested for palpable soft tissue masses and subcutaneous swellings was retrospectively collected from routine orthopedic outpatient practice. Reports were manually annotated into three triage categories: non-pathological (GREEN), non-urgent pathological (YELLOW), and urgent or potentially urgent findings (RED). A pretrained Turkish BERT model was fine-tuned for three-class classification. Model performance was evaluated using accuracy, macro-averaged F1 score, per-class precision and recall, and confusion matrices. An independent dataset of previously unseen reports was additionally used to assess robustness under real-world conditions. Results: After preprocessing and deduplication, 394 unique report segments were included. The baseline BERT model achieved an accuracy of 92.5% and a macro-averaged F1 score of 0.9106 on the test set. High classification performance was observed across all classes, with particularly reliable detection of RED reports representing urgent clinical conditions. External evaluation on independent reports demonstrated high agreement with physician annotations, with discrepancies mainly occurring in borderline or indeterminate cases. Conclusions: This study demonstrates that NLP-based analysis of superficial soft tissue ultrasound reports can effectively support urgency assessment in orthopedic practice. The proposed approach offers a practical, scalable, and image-independent solution for triage, with potential to improve workflow efficiency and facilitate timely clinical decision-making in musculoskeletal imaging. Full article

The rapid expansion of e-commerce has significantly influenced consumer purchasing behavior, making user reviews a critical source of product-related information. However, the large volume of low-quality and superficial reviews limits the ability to obtain reliable insights. This study aims to classify Turkish e-commerce reviews as either useful or useless, thereby highlighting high-quality content to support more informed consumer decisions. A dataset of 15,170 Turkish product reviews collected from major e-commerce platforms was analyzed using traditional machine learning approaches, including Support Vector Machines and Logistic Regression, and transformer-based models such as BERT and RoBERTa. In addition, a novel Multi-Transformer Fusion Framework (MTFF) was proposed by integrating BERT and RoBERTa representations through concatenation, weighted-sum, and attention-based fusion strategies. Experimental results demonstrated that the concatenation-based fusion model achieved the highest performance with an F1-score of 91.75%, outperforming all individual models. Among standalone models, Turkish BERT achieved the best performance (F1: 89.37%), while the BERT + Logistic Regression hybrid approach yielded an F1-score of 88.47%. The findings indicate that multi-transformer architectures substantially enhance classification performance, particularly for agglutinative languages such as Turkish. To improve the interpretability of the proposed framework, SHAP (SHapley Additive exPlanations) was employed to analyze feature contributions and provide transparent explanations for model predictions, revealing that the model primarily relies on experience-oriented and semantically meaningful linguistic cues. The proposed approach can support e-commerce platforms by automatically prioritizing high-quality and informative reviews, thereby improving user experience and decision-making processes. Full article

This study aims to analyze text data obtained from Turkish phone calls to manage credit risk in the banking sector and predict whether customers will fulfill their payment promises. Data cleaning was identified as a critical step to improve the quality of the texts, and various natural language processing (NLP) techniques were used. The model was built using a two-layer LSTM architecture, starting with a Self-Embedding layer, and achieved approximately 80% accuracy on the test data. The findings indicate that customers who break their payment promises often cite personal life issues such as health problems, family issues, financial difficulties, and religious beliefs to ensure reliability. These results demonstrate the importance of text data in the banking sector, the applicability of different embedding methods to Turkish datasets, and their advantages and disadvantages. Furthermore, the model built using data obtained from customer conversations can help predict credit risk more accurately and contribute to improving call center processes. Automating data cleaning processes and developing speech-to-text translation tools are recommended for future studies. Full article

The development of parallel audio corpora for Turkic languages, such as Kazakh, Uzbek, and Tatar, remains a significant challenge in the development of multilingual speech synthesis, recognition systems, and machine translation. These languages are low-resource in speech technologies, lacking sufficiently large audio datasets with aligned transcriptions that are crucial for modern recognition, synthesis, and understanding systems. This article presents the development and experimental evaluation of a speech corpus focused on Turkic languages, intended for use in speech synthesis and automatic translation tasks. The primary objective is to create parallel audio corpora using a cascade generation method, which combines artificial intelligence and text-to-speech (TTS) technologies to generate both audio and text, and to evaluate the quality and suitability of the generated data. To evaluate the quality of synthesized speech, metrics measuring naturalness, intonation, expressiveness, and linguistic adequacy were applied. As a result, a multimodal (Kazakh–Turkish, Kazakh–Tatar, Kazakh–Uzbek) corpus was created, combining high-quality natural Kazakh audio with transcription and translation, along with synthetic audio in Turkish, Tatar, and Uzbek. These corpora offer a unique resource for speech and text processing research, enabling the integration of ASR, MT, TTS, and speech-to-speech translation (STS). Full article

Irony and sarcasm are forms of expression that emphasize the inconsistency between what is said and what is meant. Correctly classifying such expressions is an important text mining problem, especially on user-centered platforms such as social media. Due to the increasing prevalence of implicit expressions, this topic has become a significant area of research in Natural Language Processing (NLP). However, the simultaneous detection of ironic and sarcastic expressions is highly challenging, as both types of implicit sentiments often convey closely related meanings. To address the detection of irony and sarcasm, this study compares the performance of transformer-based models and an ensemble learning method on Turkish texts, using five textual datasets—monogram, bigram, trigram, quadrigram, and omnigram—that share the same textual content but differ in context length. To improve classification performance, an ensemble learning approach based on the Artificial Rabbit Optimization (ARO) algorithm was implemented, combining the outputs of the models to produce final predictions. The experimental results indicate that as the context width of the datasets increases, the models achieve better predictions, leading to improvements across all performance metrics. The ensemble learning method outperformed individual models in all metrics, with performance increasing as the context expanded, achieving the highest success in the omnigram dataset with 76.71% accuracy, 74.64% precision, 73.29% sensitivity, and 73.96% F-Score. This study demonstrates that both model architecture and data structure are decisive factors in text classification performance, showing that community methods can make significant contributions to the effectiveness of deep learning solutions in low-resource languages. Full article

In the current information age, with the exponential growth of data volume and language-based applications, the accurate resolution of intra-contextual relationships in texts has become indispensable for both academic research and industrial Natural Language Processing (NLP) systems. This study focuses on the detection of pronominal anaphora in Turkish sentences. For the detection of pronominal anaphora, a specific dataset comprising 2000 sentences and 72,239 tokens was created, and this dataset was labeled using a BIO tagging method developed with a custom approach for this study. In this work, fine-tuning was performed on Transformer-based language models pre-trained on Turkish data, such as BERT and RoBERTa. Additionally, Large Language Models (LLMs) trained on Turkish data, including Turkcell-LLM-7b-v1 and ytu-ce-cosmos/Turkish-Llama-8b-DPO-v0.1, as well as multilingual models like Microsoft’s Phi-3 Mini-4K-Instruct and OpenAI’s GPT-4o-mini, were also fine-tuned with the created dataset to detect pronominal anaphora in sentences. Following the training of the language models, the resulting performance was evaluated using pronoun accuracy, antecedent accuracy, exact match, and F1-score metrics. According to the results obtained in the pronominal anaphora detection phase of the study, a novel hybrid ensemble approach combining multiple Transformer models with linguistic rules achieved the highest performance. This hybrid system attained scores of 0.987 for pronoun accuracy, 0.977 for antecedent accuracy, 0.505 for exact match, and 0.960 for F1-score, surpassing all individual models, including GPT-4o-mini. These findings reveal the superiority of ensemble methods combined with Turkish-specific linguistic rules over standalone models in Turkish anaphora resolution. This study is considered novel, as it is the first work to apply hybrid ensemble methods with linguistic rule integration to this domain for the Turkish language. Full article

Identifying age-appropriate books for children is a complex task that requires balancing linguistic, cognitive, and thematic factors. This study introduces an artificial intelligence–supported framework to estimate the Suitable Age Value (SAV) of Turkish children’s books targeting the 2–18-year age range. We employ repeated, stratified $5\times 5$ cross-validation and report out-of-fold (OOF) metrics with 95% confidence intervals for a dataset of 300 Turkish children’s books. As classical baselines, linear/ElasticNet, SVR, Random Forest (RF), and XGBoost are trained on the engineered features; we also evaluate a rule-based Ateşman readability baseline. For text, we use a frozen dbmdz/bert-base-turkish-uncased encoder inside two hybrid variants, Concat and Attention-gated, with fold-internal PCA and metadata selection; augmentation is applied only to the training folds. Finally, we probe a few-shot LLM pipeline (GPT-4o-mini) and a convex blend of RF and LLM predictions. A few-shot LLM markedly outperforms the zero-shot model, and zero-shot performance is unreliable. Among hybrids, Concat performs better than Attention-gated, yet both trail our best classical baseline. A convex RF + LLM blend, learned via bootstrap out-of-bag sampling, achieves a lower RMSE/MAE than either component and a slightly higher QWK. The Ateşman baseline performance is substantially weaker. Overall, the findings were as follows: feature-based RF remains a strong baseline, few-shot LLMs add semantic cues, blending consistently helps, and simple hybrid concatenation beats a lightweight attention gate under our small-N regime. Full article

Is negation negative? For some authors, in some languages, it is not. This is the case for so-called strict negative concord languages (e.g., Russian), in which negation is taken to be non-negative, following the cross-linguistic analysis for negative concord systems proposed by Hedde Zeijlstra’s work “Sentential negation and negative concord”. However, this analysis is focused on languages with SVO word order. In this paper, I propose to reconsider the typology of negative concord by zooming out of the focus on SVO languages that current literature has relied on. I discuss the case of SOV languages where observing a strict NC pattern leads to weaker conclusions about the nature of negation than for SVO languages with strict negative concord, leaving the negativity status of negation in those languages underdetermined. I then take a look at Turkish, an SOV language with three sentential negation markers: plain sentential negation -mA, copular negation değil, and existential negation yok. Evidence from the interaction of these markers with neither..nor phrases suggests that değil and yok, in contrast with -mA, are non-negative for some speakers. In order to explain the variation, I put forward a hypothesis about the learning process, in which there is sometimes insufficient evidence in the input to determine whether değil and yok are negative, and learners choose between two conflicting heuristics that result in the negativity or non-negativity of these markers. Full article

Objective: Named entity recognition (NER) offers a powerful method for automatically extracting key clinical information from text, but current models often lack sufficient support for non-English languages. Materials and Methods: This study investigated a prompt-based NER approach using Google’s Gemini 1.5 Pro, a large language model (LLM) with a 1.5-million-token context window. We focused on extracting important clinical entities from Turkish mammography reports, a language with limited available natural language processing (NLP) tools. Our method employed many-shot learning, incorporating 165 examples within a 26,000-token prompt derived from 75 initial reports. We tested the model on a separate set of 85 unannotated reports, concentrating on five key entities: anatomy (ANAT), impression (IMP), observation presence (OBS-P), absence (OBS-A), and uncertainty (OBS-U). Results: Our approach achieved high accuracy, with a macro-averaged F1 score of 0.99 for relaxed match and 0.84 for exact match. In relaxed matching, the model achieved F1 scores of 0.99 for ANAT, 0.99 for IMP, 1.00 for OBS-P, 1.00 for OBS-A, and 0.99 for OBS-U. For exact match, the F1 scores were 0.88 for ANAT, 0.79 for IMP, 0.78 for OBS-P, 0.94 for OBS-A, and 0.82 for OBS-U. Discussion: These results indicate that a many-shot prompt engineering approach with large language models provides an effective way to automate clinical information extraction for languages where NLP resources are less developed, and as reported in the literature, generally outperforms zero-shot, five-shot, and other few-shot methods. Conclusion: This approach has the potential to significantly improve clinical workflows and research efforts in multilingual healthcare environments. Full article