Journals / Zeki Sistemler Teori ve Uygulamaları Dergisi / 2026 / Cilt: 9 Sayı: 2026

Turkish Word Knowledge in Large Language Models: Hallucination and Abstention

Büyük Dil Modellerinin Türkçe Sözcük Bilgisi: Halüsinasyon ve Çekimserlik

Furkan Göz(KOCAELİ ÜNİVERSİTESİ)0000-0002-6726-3679
Pages
1–13
Publication Date
2026-10-28
🏛️Bibliographic metadata and full text access for this article are provided via DergiPark.
DergiPark Record ↗
✨

AI Summary & Key Takeaways

AI Insight
⚡ TL;DR(Too Long; Didn't Read)

The results indicate that Turkish vocabulary evaluations should include real rare and regional words alongside pseudowords and should report the prompt condition as a separate variable.

📌 3 Key Takeaways

🎯
Objective & Hypothesis:Erroneous generations of large language models on rare and regional vocabulary may remain invisible in evaluations built on common knowledge questions or fabricated entities.
📊
Key Findings & Data:This study examines incorrect-definition and abstention behavior in Turkish word meaning questions with a four-layer test set of 20 common, 70 rare, 70 dialectal, and 40 generated pseudowords.
💡
Conclusion & Impact:The results indicate that Turkish vocabulary evaluations should include real rare and regional words alongside pseudowords and should report the prompt condition as a separate variable.
🔬Design:Modeling & Simulation
📍Scope:Comprehensive Sample
⚙️Variables:büyük dil modelleri · halüsinasyon · çekimserlik

Abstract

Erroneous generations of large language models on rare and regional vocabulary may remain invisible in evaluations built on common knowledge questions or fabricated entities. This study examines incorrect-definition and abstention behavior in Turkish word meaning questions with a four-layer test set of 20 common, 70 rare, 70 dialectal, and 40 generated pseudowords. Nine models were evaluated under a neutral prompt asking for the meaning of a word and a prompt that explicitly allows the model to say it does not know. The resulting 3,600 responses were coded by a large language model judge (LLM judge) with a seven-category coding scheme, and the codes were validated on a blind human-coded subset of 500 responses. The primary outcome is an incorrect definition, that is, a stated meaning that does not match the dictionary definition. For this outcome, Cohen's kappa between the human coder and the judge is 0.80. According to the judge codes, the incorrect-definition rate is 1.4% for common words, 20.7% for rare words, 38.8% for dialectal words, and 26.1% for pseudowords. The rate for dialectal words is 12.7 percentage points higher than for pseudowords. The difference between rare words and pseudowords is not significant. About half of the incorrect definitions for real words are the meaning of another real word. The incorrect-definition rate is 40.2% under the neutral prompt and 12.2% under the prompt allowing abstention. No expression of uncertainty was present in 87.0% of incorrect definitions according to the judge codes and in 90.3% in the human-coded subset. One quarter of abstentions on real words assert that the word does not exist. The results indicate that Turkish vocabulary evaluations should include real rare and regional words alongside pseudowords and should report the prompt condition as a separate variable.

Özet

Büyük dil modellerinin seyrek ve bölgesel söz varlığı karşısındaki yanlış üretimleri, yaygın bilgi sorularına ya da uydurma varlıklara dayanan değerlendirmelerde yeterince görünür olmayabilir. Bu çalışma Türkçe sözcük anlamı sorularındaki yanlış tanım ve çekimserlik davranışını 20 yaygın, 70 nadir, 70 ağız ve 40 üretilmiş sahte sözcükten oluşan dört katmanlı bir test kümesiyle incelemektedir. Dokuz model, sözcük anlamını soran nötr istem ve modelin “bilmiyorum” demesine açıkça izin veren istem koşullarında değerlendirilmiştir. Elde edilen 3.600 yanıt yedi kategorili bir kodlama şemasıyla büyük dil modeli tabanlı bir yargıç (BDM-yargıç) tarafından kodlanmış, kodlar 500 yanıtlık kör insan alt kümesinde doğrulanmıştır. Ana sonuç değişkeni, sözlük tanımıyla uyuşmayan bir anlamın ileri sürülmesi olarak tanımlanan yanlış tanımdır. Bu değişken için insan ile yargıç arasındaki Cohen's kappa 0,80'dir. Yargıç kodlarına göre yanlış tanım oranı yaygın sözcüklerde %1,4, nadir sözcüklerde %20,7, ağız sözcüklerinde %38,8 ve sahte sözcüklerde %26,1'dir. Ağız sözcüklerindeki oran sahte sözcüklerdekinden 12,7 yüzde puan yüksektir. Nadir sözcükler ile sahte sözcükler arasındaki fark anlamlı değildir. Gerçek sözcüklerdeki yanlış tanımların yaklaşık yarısı başka bir gerçek sözcüğün anlamıdır. Yanlış tanım oranı nötr istemde %40,2 iken çekimserliğe izin veren istemde %12,2'dir. Yanlış tanımların yargıç kodlarına göre %87,0'ında, insan kodlu alt kümede %90,3'ünde belirsizlik ifadesi yer almamıştır. Gerçek sözcüklere verilen çekimser yanıtların dörtte biri sözcüğün var olmadığını ileri sürmektedir. Sonuçlar, Türkçe söz varlığı değerlendirmelerinin gerçek seyrek ve bölgesel sözcükleri sahte sözcüklerle birlikte içermesi ve istem koşulunun ayrı bir değişken olarak raporlanması gerektiğini göstermektedir.

🔖

Cite This Article

Göz, F. (2026). Turkish Word Knowledge in Large Language Models: Hallucination and Abstention. Zeki Sistemler Teori ve Uygulamaları Dergisi, Cilt: 9 Sayı: 2026, 1–13. https://doi.org/10.38016/jista.2027216

📄 Full Text PDF Viewer

📄

PDF görüntüleyici hazırlanıyor...

References (32)

  1. Akalın, Ş.H., 2008. Türkiye Türkçesi Ağızları Sözlüğü sanal ortamda. Türk Dili, 676, 303–315. https://agiz.tdk.gov.tr/
  2. Algayres, R., Saint-James, C.-É., Luthra, M., Shen, J., Benchekroun, Y., Lin, D., Moritz, R., Pino, J. and Dupoux, E., 2025. LongTail-Swap: Benchmarking language models' abilities on rare words. Findings of the Association for Computational Linguistics: EMNLP 2025, 11231–11251. https://doi.org/10.18653/v1/2025.findings-emnlp.601
  3. Bang, Y., Ji, Z., Schelten, A., Hartshorn, A., Fowler, T., Zhang, C., Cancedda, N. and Fung, P., 2025. HalluLens: LLM hallucination benchmark. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 24128–24156.
  4. Bayram, M.A., Fincan, A.A., Gümüş, A.S., Diri, B., Yıldırım, S. and Aytaş, Ö., 2025a. Setting standards in Turkish NLP: TR-MMLU for large language model evaluation. arXiv:2501.00593. https://doi.org/10.48550/arXiv.2501.00593
  5. Bayram, M.A., Fincan, A.A., Gümüş, A.S., Karakaş, S., Diri, B. and Yıldırım, S., 2025b. Doğal dil işlemede tokenizasyon standartları ve ölçümü: Türkçe üzerinden büyük dil modellerinin karşılaştırmalı analizi. 2025 33rd Signal Processing and Communications Applications Conference (SIU), 1–4. IEEE. https://doi.org/10.1109/SIU66497.2025.11112220
  6. Bayramoğlu, B., Fıçıcı, M., Bildirici, F., Savran Kızıltepe, R. and Karakuş, M., 2026. A white-box benchmark for intrinsic hallucination detection in Turkish LLMs. 2026 5th International Informatics and Software Engineering Conference (IISEC), 549–554. IEEE. https://doi.org/10.1109/IISEC69317.2026.11418463
  7. Chen, Z.-Y., Wang, H., Zhang, X., Hu, E. and Lin, Y., 2025. Beyond the surface: Measuring self-preference in LLM judgments. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, 1653–1672.
  8. Çöltekin, Ç., Doğruöz, A.S. and Çetinoğlu, Ö., 2023. Resources for Turkish natural language processing: A critical survey. Language Resources and Evaluation, 57(1), 449–488. https://doi.org/10.1007/s10579-022-09605-4
  9. Datta, D., Chilukuri, M.K., Kumar, Y., Ghosh, S. and Zafar, M.B., 2026. Do LLM hallucination detectors suffer from low-resource effect? Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers), 2959–2985.
  10. Er, Y.A., Kesen, I., Şahin, G.G. and Erdem, A., 2026. Cetvel: A unified benchmark for evaluating language understanding, generation and cultural capacity of LLMs for Turkish. Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers), 1052–1085. https://doi.org/10.18653/v1/2026.eacl-long.46
  11. Eray, O. and Özdinç Polat, L.N., 2026. Analysis of the performance of artificial intelligence-based language models in electrical and electronics education exam questions. Journal of Intelligent Systems: Theory and Applications, 9, 1–15. https://doi.org/10.38016/jista.1831503
  12. Guerreiro, N.M., Alves, D., Waldendorf, J., Haddow, B., Birch, A., Colombo, P. and Martins, A.F.T., 2023. Hallucinations in large multilingual translation models. Transactions of the Association for Computational Linguistics, 11, 1500–1517. https://doi.org/10.1162/tacl_a_00615
  13. Güven, Ö. and Yalnız, A., 2021. Oğuzeli (Gaziantep) ağzından Derleme Sözlüğüne katkılar. Korkut Ata Türkiyat Araştırmaları Dergisi, 6, 156–173. https://doi.org/10.51531/korkutataturkiyat.1022338
  14. hermitdave, 2018. FrequencyWords: Frequency lists by language. GitHub repository. https://github.com/hermitdave/FrequencyWords
  15. Huang, L., Yu, W., Ma, W., Zhong, W., Feng, Z., Wang, H., Chen, Q., Peng, W., Feng, X., Qin, B. and Liu, T., 2025. A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions. ACM Transactions on Information Systems, 43(2), 1–55. https://doi.org/10.1145/3703155
  16. Ji, Z., Lee, N., Frieske, R., Yu, T., Su, D., Xu, Y., Ishii, E., Bang, Y., Madotto, A. and Fung, P., 2023. Survey of hallucination in natural language generation. ACM Computing Surveys, 55(12), 1–38. https://doi.org/10.1145/3571730
  17. Kalai, A.T., Nachum, O., Vempala, S.S. and Zhang, E., 2025. Why language models hallucinate. arXiv:2509.04664. https://doi.org/10.48550/arXiv.2509.04664
  18. Kirichenko, P., Ibrahim, M., Chaudhuri, K. and Bell, S.J., 2025. AbstentionBench: Reasoning LLMs fail on unanswerable questions. Advances in Neural Information Processing Systems, 38, Datasets and Benchmarks Track, 190800–190850. https://doi.org/10.52202/085713-5729
  19. Koruyan, K., 2025. Restoran müşteri yorumlarının duygu analizi: Sıfır-atış metin sınıflandırma yaklaşımı. Journal of Intelligent Systems: Theory and Applications, 8(1), 47–62. https://doi.org/10.38016/jista.1543187
  20. Li, J., Cheng, X., Zhao, W.X., Nie, J.-Y. and Wen, J.-R., 2023. HaluEval: A large-scale hallucination evaluation benchmark for large language models. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, 6449–6464. https://doi.org/10.18653/v1/2023.emnlp-main.397
  21. Lin, S., Hilton, J. and Evans, O., 2022. TruthfulQA: Measuring how models mimic human falsehoods. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 3214–3252. https://doi.org/10.18653/v1/2022.acl-long.229
  22. Ling, Z., Liu, S., Tang, Y., Yang, J., Fu, S., Son, S., Huang, C., Huang, K., Wan, Y., Hou, Z. and Hu, X., 2025. LLM abstention can be a prompt artifact, in addition to genuine uncertainty. arXiv:2507.16199. https://doi.org/10.48550/arXiv.2507.16199
  23. Luden, I., Giulianelli, M. and Fernández, R., 2024. Beyond perplexity: Examining temporal generalization in large language models via definition generation. Computational Linguistics in the Netherlands Journal, 13, 205–232.
  24. Ruhi, Ş. and Işık Taş, E.E., 2012. Constructing general and dialectal corpora for language variation research: Two case studies from Turkish. Proceedings of the LREC 2012 Workshop on Best Practices for Speech Corpora in Linguistic Research, 17–22. http://www.lrec-conf.org/proceedings/lrec2012/workshops/03.Speech%20Corpora%20Proceedings.pdf
  25. Sun, Y., Yin, Z., Guo, Q., Wu, J., Qiu, X. and Zhao, H., 2024. Benchmarking hallucination in large language models based on unanswerable math word problem. Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), 2178–2188. https://aclanthology.org/2024.lrec-main.196/
  26. Taş, S., El Huseyni, M., Ezerceli, Ö., Bayraktar, R. and Terzioğlu, F.B., 2025. Turk-LettuceDetect: A hallucination detection models for Turkish RAG applications. arXiv:2509.17671. https://doi.org/10.48550/arXiv.2509.17671
  27. Tekin, F. and Cantürk, S., 2014. Giresun ve yöresi ağızlarından Derleme Sözlüğü'ne katkılar. Diyalektolog, 9, 33–46. https://agiz.tdk.gov.tr/tdk_arsiv_pdf/TDKAA2076c0e9-c0ca-41f5-8718-c0ccb88b68d7.pdf
  28. Toraman, Ç., Sever, A.K., Cengiz, A.A., Arslan, E.E., Sevinç, G., Birdal, M.M., Güldemir, Y.F., Kanburoğlu, A.B., Felekoğlu, S., Gürlek, O., Kantar, S., Şahin Kütük, B., Tufan, B., Genç, E., Coşkun, S., Demir, G.E., Arayıcı, M.E., Dursun, O., Gungor, O., Üsküdarlı, S., Topraksoy, A. and Darıcı, E., 2026. TurkBench: A benchmark for evaluating Turkish large language models. Proceedings of the Second Workshop on Natural Language Processing for Turkic Languages (SIGTURK 2026), 126–154.
  29. Türk Dil Kurumu, t.y.-a. Güncel Türkçe Sözlük. https://sozluk.gov.tr/
  30. Türk Dil Kurumu, t.y.-b. Derleme Sözlüğü. https://sozluk.gov.tr/
  31. Uluoğlakcı, C. and Temizel, T., 2024. HypoTermQA: Hypothetical terms dataset for benchmarking hallucination tendency of LLMs. Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics: Student Research Workshop, 95–136. https://doi.org/10.18653/v1/2024.eacl-srw.9
  32. Yüksel, A., Köksal, A., Senel, L.K., Korhonen, A. and Schuetze, H., 2024. TurkishMMLU: Measuring massive multitask language understanding in Turkish. Findings of the Association for Computational Linguistics: EMNLP 2024, 7035–7055. https://doi.org/10.18653/v1/2024.findings-emnlp.413