--- title: "NU_Internship team at ImageEval 2025: From Zero-Shot to" description: "The paper looks at Arabic image captioning in a narrow and sensitive domain: photos… 2.57 accuracy. 10 references, cited by 1 paper here." authors: ["Rana Gaber", "Seif Eldin Amgad", "Ahmed Sherif Nasri", "Mohamed Ibrahim Ragab", "Ensaf Hussein Mohamed"] year: 2025 venue: "Proceedings of The Third Arabic Natural Language Processing Conference: Shared Tasks" source: "https://aclanthology.org/2025.arabicnlp-sharedtasks.58/" doi: "10.18653/v1/2025.arabicnlp-sharedtasks.58" licence: "CC BY 4.0 (ACL Anthology)" licence_url: "https://creativecommons.org/licenses/by/4.0/" summary_licence: "Elamur Cosmic License" canonical: "https://research.elamur.ai/free/en/papers/gaber-2025-nu-internship-team-imageeval-2025/" --- [Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/gaber-2025-nu-internship-team-imageeval-2025/) [home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2025](https://md.research.elamur.ai/free/en/papers/2025/page/14/) / NU\_Internship team at ImageEval 2025: From Zero-Shot to Ensembles: Enhancing Grounded Arabic Image Captioning # NU\_Internship team at ImageEval 2025: From Zero-Shot to Ensembles: Enhancing Grounded Arabic Image Captioning **Authors:** [Rana Gaber](https://md.research.elamur.ai/free/en/authors/rana-gaber/), Seif Eldin Amgad, Ahmed Sherif Nasri, Mohamed Ibrahim Ragab and [Ensaf Hussein Mohamed](https://md.research.elamur.ai/free/en/authors/ensaf-hussein-mohamed/) · **Year:** 2025 · **Venue:** [Proceedings of The Third Arabic Natural Language Processing Conference: Shared Tasks](https://md.research.elamur.ai/free/en/venues/arabicnlp/) ([ArabicNLP 2025](https://md.research.elamur.ai/free/en/venues/arabicnlp/2025/)) · **Publisher:** [Association for Computational Linguistics](https://md.research.elamur.ai/free/en/publishers/association-for-computational-linguistics/) · **Licence:** [CC BY 4.0 (ACL Anthology)](https://creativecommons.org/licenses/by/4.0/) **Links:** [Publisher's page](https://aclanthology.org/2025.arabicnlp-sharedtasks.58/) · [PDF](https://aclanthology.org/2025.arabicnlp-sharedtasks.58.pdf) · [DOI 10.18653/v1/2025.arabicnlp-sharedtasks.58](https://doi.org/10.18653/v1/2025.arabicnlp-sharedtasks.58) · [OpenAlex](https://openalex.org/W4416037775) **Citations:** 1 citations (OpenAlex, 2026-10-14) · [1](https://md.research.elamur.ai/free/en/papers/gaber-2025-nu-internship-team-imageeval-2025/cited-by/) in our library, of the 149,439 papers whose bibliographies we read ## Summary The paper looks at Arabic image captioning in a narrow and sensitive domain: photos related to the Palestinian Nakba, from the ImageEval 2025 shared task. Captions there must be fluent Arabic and must also stay historically faithful. The team compared three vision–language models, AIN, a BLIP variant tuned on Arabic Flickr8k, and Qwen 2.5-VL, under four setups: - zero-shot with a retrieval-based rewriting step, - LoRA fine-tuning, - fine-tuning plus that rewriting step, - an LLM stacking ensemble that merges captions from the strongest systems. Training captions were normalized with Camel Tools and expanded with Gemini-generated paraphrases. Their official entry, fine-tuned BLIP with retrieval refinement, placed 5th. Later experiments showed the ensemble gave the best overall balance on semantic metrics. The value of retrieval depended on the base model. The work suggests that cheap adaptation on free-tier compute can noticeably lift captioning quality for an under-resourced language. ## Key points - The dataset has 3,471 human-captioned images: 2,718 for training and 753 for testing. Gemini-2.5-flash produced two meaning-preserving paraphrases per caption, giving multi-reference training data. - The official submission was fine-tuned BLIP + RAG, which ranked 5th. It scored cosine similarity 55.32, LLM-as-a-judge (GPT-4o) 24.87 and BLEU-1 22.77, the highest BLEU among all their configurations. - The stacking ensemble fused zero-shot AIN, fine-tuned Qwen and fine-tuned BLIP+RAG through Gemini. It reached the top LLM-as-a-judge score of 32.92 and cosine 59.17, close to zero-shot AIN's 59.69, but only 8.47 on BLEU-1. - Fine-tuning helped BLIP and Qwen but hurt AIN. The authors think AIN's broad pretraining adapted poorly to the narrow domain. - Human evaluation covered 5% of the test set on a 1–4 scale: | Criterion | Score | |---|---| | Conciseness | 2.97 | | Cultural relevance | 2.57 | | Accuracy | 2.23 | | Completeness | 2.13 | ## Interesting ideas - After generation, the system looks up the most similar gold examples from the domain and has a model rewrite its output to match their vocabulary and tone, which repairs style and obvious errors without retraining. - Candidate outputs from several systems, including one used without any tuning, are handed to an LLM meta-learner that composes a single answer, so each system's strengths can compensate for the others' weaknesses. - To build paraphrase-based multi-reference data, the authors tested several prompts on a small sample and kept only the one that best preserved meaning before scaling up. - The results show n-gram scores staying very low while semantic and judge scores were moderate to high, so a single lexical metric would have hidden real quality differences. - The same retrieval refinement helped weaker models much more than stronger ones, sometimes even lowering a strong model's scores, so any add-on layer needs to be measured per base system. ## For Mova How these ideas could serve Mova, our small open language model — its files, stages and tests: - **I1** (back): the text generator could retrieve corpus sentences realizing the same frame and adopt their wording, while logging which example motivated each change so the realization stays explainable. - **I2** (learn): when witness parsers disagree on silver data, an LLM teacher tier could see all candidate trees and propose a merged one, instead of discarding every non-unanimous sentence. - **I3** (answer): paraphrased variants of school test questions, produced with a pilot-validated prompt, would check whether readers answer from meaning rather than surface wording. - **I4** (back): the text → graph → text compressor should be scored both by surface overlap and by whether the regenerated text yields an equivalent meaning graph. - **I5** (learn): every self-induced repair rule should be evaluated separately on each parser (perceptron vs gparse), with a negative control, before it is enabled globally. *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 - [Alexandria University](https://md.research.elamur.ai/free/en/universities/alexandria-university/) — University, Alexandria, Egypt · 42 works · 287 citations - [Nile University](https://md.research.elamur.ai/free/en/universities/nile-university/) — University, Giza, Egypt · 14 works · 158 citations - [Ain Shams University](https://md.research.elamur.ai/free/en/universities/ain-shams-university/) — University, Cairo, Egypt · 13 works · 109 citations - [Nile University](https://md.research.elamur.ai/free/en/universities/nile-university-ug/) — University, Arua, Uganda · 4 works · 4 citations 📚 Shelves (5): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 18,614 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 17,965 works · [🎭 Arts and Humanities](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/) · 17,450 works · [🗣️ Language and Linguistics](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/language-and-linguistics/) · 17,296 works · [👁️ Computer Vision and Pattern Recognition](https://md.research.elamur.ai/free/en/shelves/computer-science/computer-vision-and-pattern-recognition/) · 3,486 works 🧭 Topics (2): [Semitic languages](https://md.research.elamur.ai/free/en/topics/semitic-languages/) · 3,066 works · [🖼️ Vision & language](https://md.research.elamur.ai/free/en/topics/multimodal/) · 811 works ## Cite Cite as: Rana Gaber, Seif Eldin Amgad, Ahmed Sherif Nasri, Mohamed Ibrahim Ragab and Ensaf Hussein Mohamed. 2025. NU\_Internship team at ImageEval 2025: From Zero-Shot to Ensembles: Enhancing Grounded Arabic Image Captioning. Proceedings of The Third Arabic Natural Language Processing Conference: Shared Tasks. The paper itself, not this page. BibTeX from the library's record of the work. ```bibtex @inproceedings{gaber2025nu, title = {{NU\_Internship team at ImageEval 2025: From Zero-Shot to Ensembles: Enhancing Grounded Arabic Image Captioning}}, author = {Rana Gaber, Seif Eldin Amgad, Ahmed Sherif Nasri, Mohamed Ibrahim Ragab and Ensaf Hussein Mohamed}, booktitle = {Proceedings of The Third Arabic Natural Language Processing Conference: Shared Tasks}, year = {2025}, doi = {10.18653/v1/2025.arabicnlp-sharedtasks.58}, url = {https://doi.org/10.18653/v1/2025.arabicnlp-sharedtasks.58} } ``` - [← previous Noor at BAREC Shared Task 2025: A Hybrid Transformer-Feature Architecture for Sentence-level Readability Assessment](https://md.research.elamur.ai/free/en/papers/rabih-2025-noor-barec-shared-task-2025/) - [next → Octopus: Towards Building the Arabic Speech LLM Suite](https://md.research.elamur.ai/free/en/papers/althubaiti-2025-octopus-building-arabic-speech-llm/) - [↑ back to papers of 2025](https://md.research.elamur.ai/free/en/papers/2025/page/14/) ## References 10 references, 7 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet. 1. Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, Humen Zhong, Yuanzhi Zhu, Mingkun Yang, Zhaohai Li, Jianqiang Wan, Pengfei Wang, Wei Ding, Zheren Fu, Yiheng Xu, and 8 others. 2025. [Qwen2.5-vl technical report](https://md.research.elamur.ai/free/en/papers/bai-2025-qwen2-5-vl-technical/). Preprint, arXiv:2502.13923. Ahlam Bashiti, Alaa Aljabari, Hadi Hamoud, Md. [↗](https://arxiv.org/abs/2502.13923) 2. Rafiul Biswas, Bilal Shalash, Mustafa Jarrar, Fadi Zaraket, George Mikros, Ehsaneddin Asgari, and Wajdi Zaghouani. 2025. [ImageEval 2025: The First Arabic Image Captioning Shared Task](https://md.research.elamur.ai/free/en/papers/bashiti-2025-imageeval-2025-first-arabic-image/). In Proceedings of the Third Arabic Natural Language Processing Conference (ArabicNLP 2025), Suzhou, China. Association for Computational Linguistics. the 40th Annual Meeting of the Association for Computational Linguistics, pages 311–318, Philadelphia, Pennsylvania, USA. Association for Computational Linguistics. 3. Simone Bianco, Luigi Celona, Marco Donzella, and Paolo Napoletano. 2023. Improving image captioning descriptiveness by ranking and llm-based fusion. Preprint, arXiv:2306.11593. [↗](https://arxiv.org/abs/2306.11593) 4. Rita Ramos, Desmond Elliott, and Bruno Martins. 2023. Retrieval-augmented image captioning. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 3666–3681, Dubrovnik, Croatia. Association for Computational Linguistics. [find](https://scholar.google.com/scholar?q=Rita+Ramos%2C+Desmond+Elliott%2C+and+Bruno+Martins%2E+2023%2E+Retrieval%2Daugmented+image+captioning%2E+In+Proceedings+of+the+17th+Conference+of+the+European+Chapter+of+the+Association+for+Computational+Linguisti) 5. Gheorghe Comanici, Eric Bieber, Mike Schaekermann, Ice Pasupat, Noveen Sachdeva, Inderjit Dhillon, Marcel Blistein, Ori Ram, Dan Zhang, Evan Rosen, Luke Marris, Sam Petulla, Colin Gaffney, Asaf Aharoni, Nathan Lintz, Tiago Cardal Pais, Henrik Jacobsson, Idan Szpektor, Nan-Jiang Jiang, and 3290 others. 2025. [Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities](https://md.research.elamur.ai/free/en/papers/comanici-2025-gemini-2-5-pushing-frontier/). Preprint, arXiv:2507.06261. [↗](https://arxiv.org/abs/2507.06261) 6. Ahmed Heakl, Sara Ghaboura, Omkar Thawkar, Fahad Shahbaz Khan, Hisham Cholakkal, Rao Muhammad Anwer, and Salman Khan. 2025. [Ain: The arabic inclusive large multimodal model](https://md.research.elamur.ai/free/en/papers/heakl-2025-ain-arabic-inclusive-large-multimodal/). Preprint, arXiv:2502.00094. [↗](https://arxiv.org/abs/2502.00094) 7. Hui Wei, Shenghua He, Tian Xia, Andy Wong, Jingyang Lin, and Mei Han. 2024. Systematic evaluation of llm-as-a-judge in llm alignment tasks: Explainable metrics and diverse prompt templates. ArXiv, abs/2408.13006. [↗](https://arxiv.org/abs/2408.13006) 8. Muhy Eddin Za’ter and Bashar Talafha. 2022. Benchmarking and improving arabic automatic image captioning through the use of multi-task learning paradigm. Preprint, arXiv:2202.05474. A Prompt Templates A.1 Prompt to Expand Training Dataset Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2021. [Lora: Low-rank adaptation of large language models](https://md.research.elamur.ai/free/en/papers/hu-2021-lora-low-rank-adaptation-large/). Preprint, arXiv:2106.09685. ﺍ، ،ﺓﺭ ﺍ . ﺍﺫ ﺍ Junnan Li, Dongxu Li, Caiming Xiong, and Steven C. H. Hoi. 2022. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning. A.2 Prompt used to Train our models Abdelrahman Mohamed, Fakhraddin Alwajih, El Moatez Billah Nagoudi, Alcides Inciarte, and Muhammad Abdul-Mageed. 2023. Violet: A vision-language model for Arabic image captioning with gemini decoder. In Proceedings of ArabicNLP 2023, pages 1–11, Singapore (Hybrid). Association for Computational Linguistics. [↗](https://arxiv.org/abs/2202.05474) 9. Ossama Obeid, Nasser Zalmout, Salam Khalifa, Dima Taji, Mai Oudah, Bashar Alhafni, Go Inoue, Fadhl Eryani, Alexander Erdmann, and Nizar Habash. 2020. [CAMeL tools: An open source python toolkit for Arabic natural language processing](https://md.research.elamur.ai/free/en/papers/obeid-2020-camel-tools-open-source-python/). In Proceedings of the Twelfth Language Resources and Evaluation Conference, pages 7022–7032, Marseille, France. European Language Resources Association. 10. OpenAI, :, Aaron Hurst, Adam Lerer, Adam P. Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, Aleksander Mądry, Alex Baker-Whitcomb, Alex Beutel, Alex Borzunov, Alex Carney, Alex Chow, Alex Kirillov, and 401 others. 2024. [Gpt-4o system card](https://md.research.elamur.ai/free/en/papers/openai-2024-gpt-4o-system-card/). Preprint, arXiv:2410.21276. ـ ﻭ «orig» :ﺓﺭ ﺍ ﺍ ﺍ ﺍ ﺁ ﻩ [↗](https://arxiv.org/abs/2410.21276) *Updated 31 October 2026*