GIST (Glossary of Multilingual AI Scientific Terminology).pptx

Artificial Intelligence has become one of the fastest-growing fields in recent decades, producing thousands of research papers every year. While machine translation has improved considerably and can now handle many general domains with relative ease, scientific and technical translation still faces persistent difficulties. One of the biggest challenges is the accurate and consistent translation of specialized terminology.

In AI research, technical terms like coreference resolution, explain-away effect, or attention mechanism carry precise meanings that are often lost in general-purpose translation systems such as Google Translate or DeepL. These systems tend to produce literal or contextually inaccurate renderings, leading to misunderstandings and even misinterpretations. For researchers and students who rely on translations to access cutting-edge research, these mistakes can create significant barriers.
This challenge is not only linguistic but also social and academic: mistranslated terminology reduces global inclusivity in science, since a large portion of researchers worldwide are non-native English speakers. Without standardized, high-quality translations of AI terms, many communities remain excluded from participating fully in the advancement of AI.

To respond to this gap, researchers developed GIST (Glossary of Multilingual AI Scientific Terminology). GIST is the first large-scale multilingual dataset focused specifically on AI terminology. It provides over 5,000 AI-specific terms, carefully extracted from award-winning papers between 2000 and 2023, and translated into five languages: Arabic, Chinese, French, Japanese, and Russian. The aim is to make AI research more accessible and to build a foundation for standardized scientific terminology across languages.

1. Towards Global AI Inclusivity: A Large-Scale Multilingual Terminology Dataset (GIST) An Article by Jiarui Liu et al. from Planck Institute, Vector Institute and Michigan University Reviewed and presented by Amani BOUDADA 2025/2026

2. Artificial Intelligence has become one of the fastest-growing fields in recent decades, producing thousands of research papers every year. While machine translation has improved considerably and can now handle many general domains with relative ease, scientific and technical translation still faces persistent difficulties. One of the biggest challenges is the accurate and consistent translation of specialized terminology. In AI research, technical terms like coreference resolution, explain-away effect, or attention mechanism carry precise meanings that are often lost in general-purpose translation systems such as Google Translate or DeepL . These systems tend to produce literal or contextually inaccurate renderings, leading to misunderstandings and even misinterpretations. For researchers and students who rely on translations to access cutting-edge research, these mistakes can create significant barriers. This challenge is not only linguistic but also social and academic: mistranslated terminology reduces global inclusivity in science, since a large portion of researchers worldwide are non-native English speakers. Without standardized, high-quality translations of AI terms, many communities remain excluded from participating fully in the advancement of AI. To respond to this gap, researchers developed GIST (Glossary of Multilingual AI Scientific Terminology). GIST is the first large-scale multilingual dataset focused specifically on AI terminology. It provides over 5,000 AI-specific terms, carefully extracted from award-winning papers between 2000 and 2023, and translated into five languages: Arabic, Chinese, French, Japanese, and Russian. The aim is to make AI research more accessible and to build a foundation for standardized scientific terminology across languages. Introduction 1

3. Data Abundance Computing Power General-purpose translation tools (Google Translate, DeepL , etc.) often mistranslate these, leading to loss of precision or even misinterpretation. Non-English speakers, especially researchers, face barriers to access when terminology is inconsistent or wrong. T he Critical Problem is that Technical Terminology is a Translation Nightmare Machine translation has improved dramatically in recent years, but domain-specific texts, especially in scientific and technical fields, remain challenging. Research papers contain highly specialized terms, and traditional translation cannot no longer provide accurate rendering to technical terminology. “This is not just a technical issue; it’s a matter of inclusivity and fairness in global research. If key terms are mistranslated, entire communities are excluded from understanding AI research.” 2

4. What is GIST? Precision Accessibility Adaptability 1 3 2 Ensures both coverage (through automation) and accuracy (through human review). The full dataset and code are publicly available. Integrated into platforms like ACL Anthology, allowing users to read papers with improved terminology translations. GIST = Glossary of Multilingual AI Scientific Terminology. Largest multilingual dataset of AI terminology to date. Contains 5,000+ specialized AI terms, collected from award-winning papers between 2000–2023. Languages included: Arabic, Chinese, French, Japanese, Russian. 3

5. Methodology Sources: The dataset was built from 879 awarded AI papers (Best Paper, Outstanding Paper, etc.) across 18 major specialized conferences from 2000–2023. Extraction: Researchers used LLaMA-3 and later GPT-4o to automatically identify potential technical terms. Criteria: - Terms must be nouns or noun phrases. - They must be specific to specialized fields (models , algorithms, effects, evaluation metrics, etc.). - Terms with general meanings were excluded. Filtering: Non-noun phrases, abbreviations, duplicates, and irrelevant terms were removed. Finally, three domain experts manually reviewed the list. Supplementation: Existing terminology lists from initiatives such as ACL, government glossaries, Wikipedia, and national AI dictionaries were integrated. 1. Terminology Collection 4

6. 2. Translation Process The translation process was hybrid, involving both humans and AI: Initial translations were generated using LLMs (GPT-3.5, Claude, Google Translate). Results showed low agreement (only ~15% across models), confirming the need for human input. Crowd-sourced translation: Bilingual annotators with AI expertise on Amazon Mechanical Turk translated each term. Each term received 10 independent translations per language. Annotators were carefully screened using qualification tests. Low-quality contributors were filtered out. Selection of best translation: Instead of majority voting, the authors used GPT-4o to pick the most accurate candidate among the submissions. Final human verification was conducted by experts to ensure accuracy and consistency. 5

7. 3.Integration GIST into Translation Processes using AI Prompting refinement: LLMs revise an initial translation using GIST terms. Word alignment substitution: AI terms in translations are replaced with GIST terms using multilingual smart word-matching. Constrained decoding: MT models are forced to follow GIST terminology during translation. Findings: The authors tested three post-translation refinement methods and found that Prompting method gave the most consistent improvements across languages. Word alignment worked well for Chinese and Japanese but introduced errors in morphologically rich languages like Arabic and Russian. 6

8. 7 Total: ~5,000 AI terms across five target languages. Statistics (examples): Arabic: 4,844 terms. French: 6,527 terms. Russian: 5,167 terms. Average English term length: ~2 words. Morphological differences: Chinese/Japanese → fewer inflections. Arabic/French/Russian → complex inflection systems. 👉 This diversity tested the adaptability of the framework across both high- and low-resource languages. Dataset Statistics:

9. Comparison 1: GPT-4o vs. Majority Voting GPT-4o’s candidate selection consistently produced better translations than simple majority vote. Comparison 2 : GIST vs. ACL 60-60 Dataset (which is an initiative attempt to translation about 250 technical terms into 60 languages) GIST significantly outperformed 60-60 across all five languages. Annotators judged GIST translations more accurate, consistent, and natural. Machine Translation Integration When GIST terminology was injected translation machines like AI models like Google Translate or DeepL for example, it improved the quality of their translation. 👉 Key point: Using GIST doesn’t require retraining translation models—it can be applied as a post-processing step, making it practical for real-world systems . It can also be integrated into models to enhance overall translation General Evaluation about GIST 8

10. 3. It is the f irst large-scale AI terminology dataset across five languages. Contributions: GIST makes several important contributions: 2. It c ombines LLMs with human expertise to achieve high-quality translations. 1. GIST o ffers a practical tool that improves existing MT pipelines without retraining. 4. Promotes global inclusivity in AI research. 9

11. A website demo that offers immediate translation was created on the ACL Anthology platform Researchers can now view papers with: Standard MT translation. GIST-enhanced translation (highlighting corrected terms). Direct benefit: non-English researchers gain better access to cutting-edge AI papers with accurate human-reviewed technical translation. 👉 This is a real-world proof of the concept , showing that terminology databases like GIST can be immediately useful . Practical Application 10

12. S ome terms may have multiple valid translations. Language scope limited to 5 major languages ( other languages were excluded for now). Not exhaustive: AI field evolves rapidly, new terms appear constantly. Updating requires ongoing human involvement for accuracy. Limitations 11

13. The GIST project marks an important step in bridging the gap between machine translation and specialized scientific translation. By combining the efficiency of large language models with the accuracy of human expertise, it provides a large-scale, multilingual dataset that enhances both the accessibility and the precision of AI translation. Its integration into platforms such as the ACL Anthology demonstrates real-world impact, showing how terminology consistency can empower non-English researchers to engage with cutting-edge work. While the dataset is limited to five languages and requires continual updating, GIST establishes a scalable framework for future expansion. Ultimately, it contributes to the democratization of AI knowledge, ensuring that access to scientific innovation is not limited by language barriers. Conclusion 12

14. The website Downloading Interface: 11 13

15. This is how the list appears: 14

16. Thank you for you attention. [email protected] 15 References: Liu, J., Ouzzani , I., Li, W., Zhang, L., Ou, T., Bouamor , H., Jin, Z., & Diab , M. (2025). Towards Global AI Inclusivity: A Large-Scale Multilingual Terminology Dataset (GIST). arXiv preprint arXiv:2412.18367v6 . https://huggingface.co/datasets/Jerry999/multilingual-terminology