--- title: "SAINT: Multilingual Span-Level Interpretability — Yimam 2026" description: "Sentiment classifiers are widely deployed, but their decisions are rarely checked for explainability, especially in under-resourced languages." authors: ["Seid Muhie Yimam", "Tadesse Destaw Belay", "Robert Geislinger", "Shamsuddeen Hassan Muhammad", "Adaeze Ngozi Ohuoba", "Sukairaj Hafiz Imam", "Abinew Ali Ayele", "Martin Semmann", "Chris Biemann", "Serge Sharoff"] year: 2026 venue: "Proceedings of the SIGUL 2026 Joint Workshop with ELE, EURALI, and DCLRL: Towards Inclusivity and Equality: Language Resources and Technologies for Under-Resourced and Endangered Languages" source: "https://aclanthology.org/2026.sigul-1.6/" doi: "10.63317/2ph93cb2qi57" licence: "CC BY-NC 4.0 (ELRA, LREC proceedings)" licence_url: "https://creativecommons.org/licenses/by-nc/4.0/" summary_licence: "Elamur Cosmic License" canonical: "https://research.elamur.ai/free/en/papers/yimam-2026-saint-multilingual-span-level-interpretability/" --- [Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/yimam-2026-saint-multilingual-span-level-interpretability/) [home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2026](https://md.research.elamur.ai/free/en/papers/2026/page/8/) / SAINT: Multilingual Span-Level Interpretability for Sentiment Analysis # SAINT: Multilingual Span-Level Interpretability for Sentiment Analysis **Authors:** [Seid Muhie Yimam](https://md.research.elamur.ai/free/en/authors/seid-muhie-yimam/), [Tadesse Destaw Belay](https://md.research.elamur.ai/free/en/authors/tadesse-destaw-belay/), [Robert Geislinger](https://md.research.elamur.ai/free/en/authors/robert-geislinger/), [Shamsuddeen Hassan Muhammad](https://md.research.elamur.ai/free/en/authors/shamsuddeen-hassan-muhammad/), Adaeze Ngozi Ohuoba, [Sukairaj Hafiz Imam](https://md.research.elamur.ai/free/en/authors/sukairaj-hafiz-imam/), [Abinew Ali Ayele](https://md.research.elamur.ai/free/en/authors/abinew-ali-ayele/), [Martin Semmann](https://md.research.elamur.ai/free/en/authors/martin-semmann/), [Chris Biemann](https://md.research.elamur.ai/free/en/authors/chris-biemann/), [Serge Sharoff](https://md.research.elamur.ai/free/en/authors/serge-sharoff/) · **Year:** 2026 · **Venue:** Proceedings of the SIGUL 2026 Joint Workshop with ELE, EURALI, and DCLRL: Towards Inclusivity and Equality: Language Resources and Technologies for Under-Resourced and Endangered Languages · **Licence:** [CC BY-NC 4.0 (ELRA, LREC proceedings)](https://creativecommons.org/licenses/by-nc/4.0/) **Links:** [Publisher's page](https://aclanthology.org/2026.sigul-1.6/) · [PDF](https://aclanthology.org/2026.sigul-1.6.pdf) · [DOI 10.63317/2ph93cb2qi57](https://doi.org/10.63317/2ph93cb2qi57) · [OpenAlex](https://openalex.org/W7166572793) **Citations:** 0 citations (openalex-snapshot · 2026-10-14) ## Summary Sentiment classifiers are widely deployed, but their decisions are rarely checked for explainability, especially in under-resourced languages. The authors build SAINT, a manually annotated dataset for Amharic, English, German and Hausa. Each text has an overall sentiment label (positive, negative or neutral) and word-level marks for the spans that carry sentiment and their targets. One trained expert per language annotated the data using an adapted Potato tool, and several co-authors reviewed the labels. The study compares fine-tuned encoder models, trained monolingually and multilingually, with zero- and few-shot prompted LLMs. Captum gradient attributions show which tokens drive encoder predictions. For span-level tagging, few-shot LLMs perform best in most languages. Hausa is the exception, where a multilingual encoder leads. For whole-text labels, encoders are generally strong, but LLMs can match or exceed them depending on the prompting strategy. The work provides an openly released benchmark for interpretable sentiment analysis beyond high-resource languages. ## Key points - SAINT sizes: Amharic 1,519, English 1,501, German 1,498 and Hausa 1,879 texts. The train/dev/test splits are 912/153/454 for Amharic, 899/151/451 for English, 899/150/449 for German and 1,127/189/563 for Hausa. - The sources differ by language. Amharic comes from pre-2023 X posts and news comments, and English from posts on anti-vaccination discourse. Hausa draws on forums, news portals and social media. - The encoders are BERT-base, BERT-base-german, XLM-RoBERTa-base, AfroXLMR-large and AfroXLMR-social. The LLMs are GPT-4o-mini, GPT-4.1-mini, gpt-oss-120b, Qwen3-32b and LlaMA-3.3-70B-it, prompted to return a JSON label, a rationale and per-word BIO-style sentiment tags. - Inter-annotator agreement is not reported. Quality control relied on one expert annotator per language plus review by the co-authors. - The dataset, annotation guideline and evaluation scripts are released at github.com/uhh-hcds/SAINT. *Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.* Fields: [Linguistics & languages](https://md.research.elamur.ai/free/en/shelves/linguistics/) ## Cite The paper itself, not this page. BibTeX from the library's record of the work. ```bibtex @inproceedings{yimam2026saint, title = {{SAINT: Multilingual Span-Level Interpretability for Sentiment Analysis}}, author = {Yimam, Seid Muhie and Belay, Tadesse Destaw and Geislinger, Robert and Muhammad, Shamsuddeen Hassan and Ohuoba, Adaeze Ngozi and Imam, Sukairaj Hafiz and Ayele, Abinew Ali and Semmann, Martin and Biemann, Chris and Sharoff, Serge}, booktitle = {Proceedings of the SIGUL 2026 Joint Workshop with ELE, EURALI, and DCLRL: Towards Inclusivity and Equality: Language Resources and Technologies for Under-Resourced and Endangered Languages}, year = {2026}, doi = {10.63317/2ph93cb2qi57}, url = {https://doi.org/10.63317/2ph93cb2qi57} } ``` - [← previous SagaQA: A Multi-hop Reasoning Benchmark for Long-form Narrative Understanding in TV Series](https://md.research.elamur.ai/free/en/papers/pennec-2026-sagaqa-multi-hop-reasoning-benchmark/) - [next → SalamahBench: Dialect and Category Level Safety Evaluation of Arabic Language Models](https://md.research.elamur.ai/free/en/papers/abdelnasser-2026-salamahbench-dialect-category-level-safety/) - [↑ back to papers of 2026](https://md.research.elamur.ai/free/en/papers/2026/page/8/) ## References 2 references, 0 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet. 1. Md Shah Alam, Md Sabbir Hossain Mrida, and Md Atikur Rahman. 2025. Sentiment analysis in social media: How data science impacts public opinion knowledge integrates natural language processing (NLP) with artificial intelligence (AI). American Journal of Scholarly Research and Innovation, 4(01):63–100… [find](https://scholar.google.com/scholar?q=Md+Shah+Alam%2C+Md+Sabbir+Hossain+Mrida%2C+and+Md+Atikur+Rahman%2E+2025%2E+Sentiment+analysis+in+social+media%3A+How+data+science+impacts+public+opinion+knowledge+integrates+natural+language+processing+%28NLP%29+wi) 2. Limitation Tadesse Destaw Belay, Israel Abebe Azime, Ibrahim Said Ahmad, David Ifeoluwa Adelani, Idris Abdulmumin, Abinew Ali Ayele, Shamsuddeen Hassan Muhammad, and Seid Muhie Yimam. 2025. AfroXLMR-Social: Adapting Pre-trained Language Models for African Languages Social Media Text. arXiv preprint… [find](https://scholar.google.com/scholar?q=Limitation+Tadesse+Destaw+Belay%2C+Israel+Abebe+Azime%2C+Ibrahim+Said+Ahmad%2C+David+Ifeoluwa+Adelani%2C+Idris+Abdulmumin%2C+Abinew+Ali+Ayele%2C+Shamsuddeen+Hassan+Muhammad%2C+and+Seid+Muhie+Yimam%2E+2025%2E+AfroXLMR%2DS)