LLMs as a Tool for Specialised Terminology: Study Compares GPT-4o, GPT-5.2, Claude Sonnet 4.5, and DeepSeek
A recent arXiv study (2607.24784) evaluates the potential of large language models (LLMs) in replacing traditional resources for specialized translation from English to French. Analyzing four proprietary models—GPT-4o, GPT-5.2, Claude Sonnet 4.5, and DeepSeek—the research focuses on the fields of Earth, Environmental, and Planetary Sciences (EEPS) as well as Natural Language Processing (NLP). The examination includes 80 terms from each area and assesses two prompting strategies: terminology mode and translation mode. Findings indicate notable performance disparities among the models, with Claude Sonnet 4.5 emerging as the most effective, highlighting variances in LLM utility for translators.
Key facts
- arXiv paper 2607.24784 examines LLMs for specialised terminology in English-to-French translation.
- Four proprietary models tested: GPT-4o, GPT-5.2, Claude Sonnet 4.5, and DeepSeek.
- Domains studied: Earth, Environmental and Planetary Sciences (EEPS) and Natural Language Processing (NLP).
- Experiment based on 80 terms per domain.
- Two prompting strategies compared: terminology mode and translation mode.
- Claude Sonnet 4.5 achieved best results in most favorable conditions.
- Clear differences found between models, prompting strategies, and domains.
- LLMs show potential but limitations remain compared to traditional corpora.
Entities
Institutions
- arXiv