This is a page that curates AI-related papers published worldwide. All content here is summarized using Google Gemini and operated on a non-profit basis. Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.
Science Across Languages: Assessing LLM Multilingual Translation of Scientific Papers
Created by
Haebom
Author
Hannah Calzi Kleidermacher, James Zou
Outline
This paper addresses the reality that most academic journals are published only in English, posing a barrier to non-native English speakers. This paper proposes an automatic translation system for academic papers using a large-scale language model (LLM). While maintaining the JATS XML format, we translated papers from various scientific fields into 28 languages and measured the translation accuracy (average 95.9%) using a unique question-answering (QA)-based evaluation method. A user study with 15 researchers confirmed the accuracy of the translations and also revealed differences in preferences for overtranslation of certain technical terms. Furthermore, we demonstrate the adaptability and usefulness of LLM-based translation by utilizing in-context learning techniques to mitigate the problem of overtranslation. The source code and translated papers are available at https://hankleid.github.io/ProjectMundo .
Takeaways, Limitations
•
Takeaways:
◦
Presenting the feasibility of an automatic translation system for academic papers using LLM.
◦
Increased applicability to real-world academic journals by maintaining the JATS XML format.
◦
Objective evaluation of translation accuracy through a question-and-answer-based evaluation method.
◦
Wide accessibility through support for multiple languages (28 languages).
◦
Improving translation quality and offering user customization possibilities through in-context learning.
•
Limitations:
◦
Excessive translation issues for some technical terms and the need for additional adjustments based on user preferences.
◦
The user study size was relatively small (15 people).
◦
Further research is needed to determine whether there is bias in specific areas.
◦
System architecture dependent on the performance of LLM.
