尽管生成式人工智能翻译工具输出的文本往往读起来自然流畅,但流利度并不等同于准确性,其背后可能潜藏着难以察觉的误译问题 1。The Conversation Australia 的分析指出,AI翻译的主要错误根源包括学习数据不足、提示词不充分以及模型产生的“幻觉”现象,即自信地呈现虚假信息 1。这些缺陷在词汇、句子及篇章层面均有体现:例如在澳大利亚语境下,AI曾将指代“人字拖”的单词“thongs”误译为内裤(G-strings),或因中英文时态表达差异导致时态错误 1。此外,由于AI是通过预测下一个词而非真正理解语境来决定含义,其在处理歧义时容易在句子层面出错;而在篇章层面,涉及英语使用定冠词指代前文而中文需重复词语等衔接性问题时,AI也可能出现连接错误 1。
除了基础的语言转换失误,AI还可能重现过时或不敏感的术语,这是因为新出现的包容性表达在其训练数据中代表性不足 1。为了提高翻译可靠性,建议确保话题熟悉且无文化障碍、原文语法完美清晰、语种为广泛使用的语言(如中英法西),并提供明确有效的提示词 1。专用神经机器翻译技术通常比通用聊天机器人能提供更准确的翻译结果 1。鉴于上述局限性,作者 Alisa Tian 建议在涉及文化差异、敏感话题或复杂语境时,使用者应保持谨慎,并结合人工审核以确保翻译质量 1。
Generative artificial intelligence translation tools often produce fluent text that appears accurate, yet they remain prone to subtle and potentially significant errors. The primary sources of these mistakes include insufficient training data, inadequate prompts, and "hallucinations," where the model confidently presents false information 1. While the output may sound natural, fluency does not equate to correctness, particularly when the technology fails to grasp deeper contextual nuances 1.
Errors manifest across multiple linguistic levels. At the lexical level, misinterpretations can occur due to cultural differences or grammatical discrepancies; for instance, the Australian term for flip-flops ("thongs") might be incorrectly translated into Chinese as underwear (G-strings), and tense distinctions between languages like Chinese and English may be mishandled 1. Sentence-level issues arise from ambiguity, as AI models predict the next word based on statistical patterns rather than truly understanding context 1. Furthermore, discourse-level problems involve cohesion; since English uses definite articles to refer back to previous mentions while Chinese often requires repeating words, AI may fail to maintain proper logical connections 1. There is also a risk that these systems will revive outdated or insensitive terminology because newer inclusive expressions are underrepresented in their training datasets 1.
To mitigate these risks, users should exercise caution when dealing with topics involving cultural sensitivity or complex contexts, ideally combining AI tools with human review 1. Reliability improves when the subject matter is familiar, free of cultural barriers, written in perfect grammar, involves widely spoken languages such as English, Chinese, French, or Spanish, and is accompanied by clear prompts 1. However, dedicated neural machine translation technologies generally provide more accurate results than general-purpose chatbots 1.
评论
还没有评论,欢迎留下第一条。