Quality assessment and post-editing of agricultural science and technology texts translated by large language models
基于大语言模型的农业科技文本翻译质量评估及译后编辑

摘要

为实现大语言模型在农业科技翻译中译文质量与效率的平衡，译后编辑成为关键的翻译实施方式。本研究以《中国玉米病虫草害图鉴》为语料，基于多维质量指标（MQM）框架，系统评估文心一言与DeepSeek在农业科技汉译英任务中的翻译质量。研究发现，二者存在共性缺陷与差异性短板。准确性错误占比最高（文心一言47.56%，DeepSeek41.92%），主要表现为形态参数、防治数据等关键信息失真。差异性方面，文心一言的准确性错误更为突出，DeepSeek则在格式与风格规范层面短板明显。针对上述问题，本文提出三大译后编辑策略：修正误译与增补漏译以提升准确性，术语溯源与知识验证以减少术语错误，释义性补充以强化区域惯例适配。研究证实，通用大语言模型在农业科技翻译中仍需人工干预，译后编辑应聚焦于专业知识的精准重构与文化语境的深度适配。

Abstract

To achieve a balance between translation quality and efficiency in agricultural science and technology translation using large language models, post-editing has emerged as a key translation implementation method. This study employs the Illustrated Handbook of Maize Diseases, Insect Pests and Weeds in China as its corpus and based on the Multidimensional Quality Metrics (MQM) framework, systematically evaluates the translation quality of ERNIE Bot and DeepSeek in Chinese-to-English agricultural science and technology translation tasks. The study reveals both common deficiencies and distinct weaknesses between the two models. Accuracy errors account for the highest proportion (47.56% for ERNIE Bot and 41.92% for DeepSeek), primarily manifesting as distortions in key information such as morphological parameters and control data. In terms of differences, ERNIE Bot exhibits more prominent accuracy errors, while DeepSeek shows more pronounced deficiencies in formatting and stylistic norms. To address these issues, this paper proposes three post-editing strategies: correcting mistranslations and supplementing omissions to enhance accuracy, tracing term origins and conducting knowledge verification to reduce terminology errors, and adding explanatory supplements to strengthen locale-convention adaptation. The study confirms that general-purpose large language models still require human intervention in agricultural science and technology translation, and that post-editing should focus on the precise reconstruction of specialized knowledge and the deep adaptation of cultural contexts.