Elamur Research⌕
☰
home / papers / 2026 / LLM-Based Financial Sentiment Analysis in Arabic: Evidence from Saudi Markets

LLM-Based Financial Sentiment Analysis in Arabic: Evidence from Saudi Markets

Paper · 2026 · Mona H. Albaqawi, Eman M. Albalkhi, Joud A. Albaiti et al.

Summary

The paper addresses fine-grained sentiment analysis of Arabic financial text for the Saudi stock market, where resources are scarce and the language is morphologically complex. The authors built a pipeline that collects official financial news and social media posts through APIs. It cleans and normalizes the Arabic text, removes duplicates at several levels, and links company mentions to canonical identifiers. A CAMeLBERT NER model is combined with a curated alias lexicon for this linking step. Long texts are summarized first, and every text is then labeled on a five-point sentiment scale by three LLMs whose votes are combined. The result is a corpus of roughly 84K samples covering all companies listed on the Saudi Exchange. In classification experiments, zero-shot GPT-5 reached Macro-F1 0.829. That is far above fine-tuned AraBERT and CAMeLBERT (about 0.54), an SVM and a sentiment lexicon. The work shows that LLM-driven annotation and classification can scale Arabic domain NLP. However, the corpus is not publicly released.

Key points

  • The corpus has about 84K samples: about 62.8K from social media (74.8%) and about 21.2K from news (25.2%). The five labels run from strongly negative to strongly positive, and all 261 TASI-listed companies are mentioned.

  • Labels come from three models: GPT-4 Turbo, GPT-4o Mini and Gemini 2.5 Flash.

    • Pairwise Cohen's κ ranges from 0.505 to 0.611.
    • At least two of three models agree on 95.16% of social posts and 96.06% of news items.
    • All three agree on 53.59% and 61.58% respectively.
    • A GPT-4o judge on 250 samples served as the gold reference.

  • Entity linking pairs CAMeLBERT NER with an alias dictionary, using fuzzy matching at a 0.80 similarity threshold. A 1,584-term financial lexicon helps with detection and filtering.

  • Texts over 100 words are summarized before labeling. News averages 268 words, while social posts average 24. ALLaM was chosen as the summarizer, with a hybrid ROUGE-L plus cosine score of 0.659 at no inference cost.

  • Macro-F1 scores:

    ModelMacro-F1
    GPT-5 (zero-shot)0.829
    CAMeLBERT (fine-tuned)0.547
    AraBERT (fine-tuned)0.524
    SVM with TF-IDF0.351
    General Arabic sentiment lexicon0.287

    Most errors fall between Neutral and Positive.

Interesting ideas

  • I1. Instead of trusting a single annotator model, the authors keep a label only when a majority of independent LLMs agree, and they separate unanimous items from majority-only items as two confidence tiers.
  • I2. Agreement is checked with more than κ: Jensen–Shannon divergence and chi-square on label distributions reveal whether models are systematically skewed even when pairwise agreement looks fine.
  • I3. Entity mentions are resolved by stacking a statistical NER with a hand-curated alias dictionary under fuzzy matching, so formal names, abbreviations and informal variants all collapse to one canonical ID.
  • I4. Inputs are routed by length, so long documents are condensed first and short ones go straight to the classifier, which keeps the downstream step's input size uniform and cheap.
  • I5. Summarizers are selected on a combined score that weighs compression, semantic similarity, ROUGE and explicit hallucination measures, rather than overlap scores alone.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

  • I1 (learn): Silver data from witness parsers or LLM teachers could be split the same way, training first on unanimous trees and treating 2-of-3 items as a weaker tier.
  • I2 (learn): Comparing the distribution of relations or frame roles across witness parsers would expose a teacher that consistently over-predicts some label, a bias that per-item agreement can miss.
  • I3 (world): Entities in the world of a text could be grounded the same way, using alias lists from Wikidata and a similarity threshold to merge surface variants into a single knowledge-graph node.
  • I4 (frames): Very long passages could first be reduced to their core clauses before frame extraction, provided the reduction step stays traceable to the source sentences.
  • I5 (back): The text → graph → text compressor could be scored the same way, penalizing generated content that introduces entities or facts absent from the meaning graph.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓University of LondonLondon · United Kingdom12 works · 924 citations🎓George Mason UniversityFairfax · United States24 works · 1,288 citations🎓Damascus UniversityDamascus · Syria4 works · 43 citations

Also on the byline: University of Jeddah.

 Fields, topics and concepts

📚 Shelves 6The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksDecision Sciences · 209 worksManagement Science and Operations Research · 108 worksEconomics, Econometrics and Finance · 17 worksFinance · 5 works

🧭 Topics 2Research subjects and languages the work is about.Sentiment & emotion · 972 worksSemitic languages · 2,908 works

Cite

Cite as Mona H. Albaqawi, Eman M. Albalkhi, Joud A. Albaiti and Enrico Lopedoto. 2026. LLM-Based Financial Sentiment Analysis in Arabic: Evidence from Saudi Markets. The 7th Workshop on Open-Source Arabic Corpora and Processing Tools (OSACT7) with 5 Shared Tasks.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.63317/3hsd9iyq472z

@inproceedings{albaqawi2026llm,
  title = {{LLM-Based Financial Sentiment Analysis in Arabic: Evidence from Saudi Markets}},
  author = {Mona H. Albaqawi, Eman M. Albalkhi, Joud A. Albaiti and Enrico Lopedoto},
  booktitle = {The 7th Workshop on Open-Source Arabic Corpora and Processing Tools (OSACT7) with 5 Shared Tasks},
  year = {2026},
  doi = {10.63317/3hsd9iyq472z},
  url = {https://doi.org/10.63317/3hsd9iyq472z}
}

References

19 references, 8 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Ahmed Abdelali, Maram Hasanain, Hamdy Mubarak, Laura Kallmeyer, Hassan Sajjad, Fahim Dalvi, et al. ALLaM: Large language models for arabic and english. arXiv preprint arXiv:2407.15390, 2024. SDAIA Arabic foundation model. ↗
  2. Hero O. Ahmad and Shahla U. Umar. Sentiment analysis of financial textual data using machine learning and deep learning models. Informatica, 47(5):153–158, 2023. ↗
  3. Aisha Alansari and Hamzah Luqman. Arahallueval: A fine-grained hallucination evaluation framework for arabic llms. arXiv preprint, 2025.
  4. Mohammed Alshahrani, Fuxi Zhu, Mohammed Alghaili, Eshrag Refaee, and Mervat Bamiah. Borsah: An arabic financial tweets corpus. In Proceedings of the International Conference on Computational Linguistics, 2018. Arabic financial sentiment analysis corpus for TADAWUL. find
  5. Martins Amola. Sentiment analysis in financial news: Enhancing predictive models for stock market behavior. Preprint, 2025. Available at ResearchGate. find
  6. Wissam Antoun, Fady Baly, and Hazem Hajj. AraBERT: Transformer-based model for arabic language understanding. In Proceedings of the 4th Workshop on Open-Source Arabic Corpora and Processing Tools (OSACT), pages 9–15. European Language Resources Association (ELRA), 2020. ↗
  7. Dogu Tan Araci. Finbert: Financial sentiment analysis with pre-trained language models. arXiv preprint, 2019. ↗
  8. Gilbert Badaro, Ramy Baly, Hazem Hajj, Nizar Habash, and Wassim El-Hajj. A light lexicon-based mobile application for sentiment mining of arabic tweets. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC), pages 1703–1709, 2018. ArSenL: Arabic Sentiment Lexicon. find
  9. Nicholas Barberis, Andrei Shleifer, and Robert Vishny. A model of investor sentiment. Journal of Financial Economics, 1998. ↗
  10. Savita Bhat and Vasudeva Varma. Large language models as annotators: A preliminary evaluation for annotating low-resource language content. In Proceedings of the 4th Workshop on Evaluation and Comparison of NLP Systems. Association for Computational Linguistics, 2023. find
  11. Kelvin Du, Frank Xing, Rui Mao, and Erik Cambria. Financial sentiment analysis: Techniques and applications. ACM Computing Surveys, 56(9):220, 2024. ↗
  12. Ismail El Bazi and Nabil Laachfoubi. Arabic named entity recognition using deep learning approach. International Journal of Electrical and Computer Engineering, 9(3):2025–2032, 2019. ↗
  13. Jiawei Gu, Xuhui Jiang, Zhichao Shi, Hexiang Tan, Xuehao Zhai, Chengjin Xu, Wei Li, Yinghan Shen, Shengjie Ma, Honghao Liu, et al. A survey on llm-as-a-judge. arXiv preprint arXiv:2411.15594, 2024. ↗
  14. Huang Huang, Fei Zhu, Jianfeng Qin, Yulei Tang, Xuebai Lin, Guo Liu, and Wei Wang. AceGPT: Localizing large language models in arabic. arXiv preprint arXiv:2309.12053, 2024. Arabic-specialized instruction-tuned model. ↗
  15. Go Inoue, Bashar Alhafni, Nurpeiis Baimukan, Houda Bouamor, and Nizar Habash. The interplay of variant, size, and task type in arabic pre-trained language models. In Proceedings of the Sixth Arabic Natural Language Processing Workshop (WANLP), pages 92–104. Association for Computational Linguistics, 2021. CAMeL-BERT model family.
  16. Neha Sengupta, Sunil Kumar Sharma, Muhammed Masoud, Abbas Akkasi, Karthik Kamur, Shivani Bhatia, Ebtesam Almazrouei, et al. Jais and jais-chat: Arabic-centric foundation and instruction-tuned open generative large language models. arXiv preprint arXiv:2308.16149, 2023. 13B parameter Arabic-centric LLM from Inception/G42. ↗
  17. Sahar Sohangir, Nicholas Petty, and Dingding Wang. Financial sentiment lexicon analysis. In 2018 12th IEEE International Conference on Semantic Computing (ICSC), pages 286– 289, 2018. find
  18. Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. Self-consistency improves chain-of-thought reasoning in language models. arXiv preprint arXiv:2203.11171, 2023. ↗
  19. Shijie Wu, Ozan Irsoy, Steven Lu, Vadim Dabravolski, Mark Dredze, Sebastian Gehrmann, Prabhanjan Kambadur, David Rosenberg, and Gideon Mann. Bloomberggpt: A large language model for finance. arXiv preprint, 2023. ↗

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/albaqawi-2026-llm-based-financial-sentiment-analysis/