Breaking
CNSmart Energy Week debuts, Yuanjing’s three major solar products are launching to attack the rooftop marketITWar Ukraine Russia, missile attack on Kiev: two dead. LIVETRFinancial branch teams are searching the Halk TV buildingITUkraine, Zelensky: "Raid on thermoelectric power plant in Kiev, allies did not warn"TRPolice raid on Halk TV: Broadcasting was stopped, it was announced that a trustee was appointedCRYPTO-FRBitcoin: CPI, banks and Beige Book, the 3 meetings of the dayTRMeteorology gave the time: Heavy rain and storm will be effective! | It will last for days: Yellow and orange rain alert in many provinces, including IstanbulCNEmail reminders for parking fees are a scam! Yuantong Telecom teaches you "2 tricks" to prevent fraudCRYPTO-ENSouth Korea locks out crypto users as 27 exchange sites go darkARThe International Monetary Fund warns of widening financial imbalances and calls for urgent reformsCNSmart Energy Week debuts, Yuanjing’s three major solar products are launching to attack the rooftop marketITWar Ukraine Russia, missile attack on Kiev: two dead. LIVETRFinancial branch teams are searching the Halk TV buildingITUkraine, Zelensky: "Raid on thermoelectric power plant in Kiev, allies did not warn"TRPolice raid on Halk TV: Broadcasting was stopped, it was announced that a trustee was appointedCRYPTO-FRBitcoin: CPI, banks and Beige Book, the 3 meetings of the dayTRMeteorology gave the time: Heavy rain and storm will be effective! | It will last for days: Yellow and orange rain alert in many provinces, including IstanbulCNEmail reminders for parking fees are a scam! Yuantong Telecom teaches you "2 tricks" to prevent fraudCRYPTO-ENSouth Korea locks out crypto users as 27 exchange sites go darkARThe International Monetary Fund warns of widening financial imbalances and calls for urgent reforms
BackAI翻譯:還有多久會完全取代人類?
AI翻譯:還有多久會完全取代人類?
Tech
BBC中文2 hours agoTech3 min readChinaView translation

AI翻譯:還有多久會完全取代人類?

Quick Look

儘管人工智能翻譯在處理諺語和日常文本時表現良好,但在法律、醫療等專業領域仍存在嚴重錯誤,如誤譯單位或法律術語,導致潛在的重大後果,專家警告即使翻譯流暢也不代表準確,人類校對仍不可或缺。

AI-generated summary

Why It Matters

人工智能翻譯工具如谷歌翻譯、DeepL等被廣泛使用,月翻譯字數達一萬億,但專業領域如法律和醫療仍顯示出嚴重錯誤,例如誤譯法律術語或醫療單位,導致潛在危害。

Font size

作者,約翰·勞倫森(John Laurenson)

角色,商業事務記者

報導來源,法國巴黎

發表於 4 分鐘前

閱讀時間: 3 分鐘

法國人在下大雨時會說「Il pleut des cordes」。

直譯是「天上在下繩子」(It's raining ropes)。這類慣用語,往往會難倒人工智能翻譯。

「你可以想像,這會令人工智能陷入怎樣的兩難。」法國國家科學研究中心(CNRS)專攻自動翻譯的研究主任弗朗索瓦·伊馮(François Yvon)說。

他表示,訓練良好的系統會知道,天上並不是真的掉下繩子,因此會使用英語中的對應說法「it's raining cats and dogs」(傾盆大雨)。

伊馮問道:「但當它翻譯成印尼語時,能找到對應的說法嗎?」

評論天氣的風險相對不高,但如今愈來愈多翻譯工作由機器完成,而涉及的內容遠比天氣更重要。

谷歌今年4月表示,人們現時透過谷歌翻譯(Google Translate),每月翻譯的字數達到一萬億。

這還未計算其他工具:DeepL、微軟(Microsoft)、Claude、ChatGPT和Gemini,更不用說為企業量身訂造的翻譯系統。

Skip 熱讀 and continue reading

熱讀

End of 熱讀

去年,機器翻譯的直接銷售額(即不包括一般人工智能訂閱服務)估計達14億美元(約10億英鎊)。

瑞士非營利組織「開放知識協會」(Open Knowledge Association, OKA)創辦人約翰·齊默爾曼(John Zimmerman)說:「在許多情境中,直接採用人工智能翻譯已是正確選擇。若你翻譯的是內部文件或產品目錄,為了糾正最後那幾個百分點的錯誤,而把成本提高100倍,並不合理。」

「但百科全書就不同了。維基百科建立在信任之上。一個錯誤的事實可能存在多年,並被記者和人工智能系統引用。」

他的組織負責把維基百科文章翻譯成其他語言,目前已發表逾萬篇文章,主要翻譯成西班牙語、葡萄牙語和英語,是向這個協作式網絡百科全書貢獻翻譯文章最多的單一機構。

而如果沒有人工智能,這一切根本不可能實現。

齊默爾曼解釋:「如果用傳統方式翻譯一萬篇文章,成本會高達數百萬。但在我們的流程中,人工智能先產出初稿,再由人類譯者逐句檢查。」

然而,人類的參與正變得愈來愈少。

齊默爾曼說:「品質如今已經好到一個程度,讓完全跳過人工翻譯成為一個認真的選項,也成為業界普遍面臨的誘惑。兩三年前,使用 DeepL 等工具時,我們的譯者仍需要重寫每篇初稿中相當大的一部分。」

「如今,很多譯文的文字表達往往根本不需要修改。因此,問題已不再是 AI 翻譯是否夠好,答案顯然是肯定的;真正的問題是,它是否已經好到可以取代人類譯者。」

不過,譯者們認為,現在還不到那一步。

阿里埃拉·巴爾蒂(Ariella Baltié)是一名翻譯員兼法律專家,專門將法院文件從英文翻譯成法文。

最近有人聯絡她,表示已使用人工智能翻譯自己的文件,希望她協助校對。

「一開頭就有兩個重大錯誤。」她說。

人工智能把法文中的「Le demandeur」(原告)翻譯成了「the requester」(請求人)。

巴爾蒂表示:「對法律文件而言,這種譯法非常不恰當。」

當人工智能看到「Madame X, épouse Y」(X 女士,Y 先生的妻子)時,卻將其翻譯成「Ms X marries Y」(X 女士嫁給 Y 先生)。

換言之,它把名詞誤當成動詞,完全扭曲了原本的意思。

法律範疇的誤譯,後果可以很嚴重。以美國為例,機器翻譯的庇護申請出現錯誤,已導致被羈留者延遲獲釋。

在醫療領域,準確性更是攸關生死。

從事醫學翻譯的卡羅爾·布勞特(Carol Bereuter)說:「藥物劑量和驗血結果的單位,往往包含希臘字母。」

她曾遇到一份自動翻譯的文本,把希臘字母μ譯成了milli(千分之一),而不是micro(百萬分之一)。

「這是嚴重的錯誤。如果本該給一微克,卻給了一毫克,病人就是服了一千倍的劑量。」

她還指出,隨著人工智能翻譯變得愈來愈好,發現錯誤反而變得更加困難。

「以前用谷歌翻譯等工具,句子會很突兀,因為會有語法錯誤,或者讀起來很生硬;如今讀起來完全合理,非常流暢、自然,所以錯誤變得更難察覺。」

專攻自動翻譯的法國學者弗朗索瓦·伊馮警告:「如果譯文要由你署名,那就涉及你的聲譽風險。如果有人根據翻譯內容來決定是否接受某項醫療程序、簽署合約,或進行一項重大買賣,風險就更大了。」

他表示,即使是人類看來相當清楚的句子,也可能難倒人工智能翻譯服務。

「當一個詞有不同含義時,人類通常很清楚哪個意思適用,因為他們理解上下文;但當機器翻譯遇到『translate into Polish』(翻譯成波蘭語)這句話時,它完全有可能認定這句話講的是鞋油(Polish有『波蘭語』和『擦亮劑』兩個意思)。」

上下文往往至關重要。如果人工智能翻譯沒有意識到我們談的是足球比賽,就可能把「taking a corner」(踢角球)譯錯。

去年,OKA進行了一項由維基媒體基金會(Wikimedia)資助的研究,涵蓋119篇文章、10組語言配對,以及逾1,000小時的譯者工作時間。

OKA創辦人齊默爾曼說:「在出版前,27% 的人工智能生成文本仍被修改。這些修改中有四分之三是糾正人工智能引入的錯誤,而其中 5% 至 6% 的錯誤改變了原文的意思。」

要徹底消除這些錯誤,可能相當困難。

原因在於,人工智能是以概率為基礎,決定譯出哪個詞而非另一個詞;而「很可能正確」並不等於「真的正確」,尤其是在語言這樣細膩而充滿變化的領域中,更是如此。

他認為,新一代技術將帶來改善。

What to Watch

AI outlook — possibilities, not facts

  • 人工智能翻譯在專業領域將仍需要人類校對作為質量控制步驟

    Likely · Within months

  • 開放知識協會等機構將繼續使用人工智能生成初稿,再由人類譯者逐句檢查的混合模式

    Very likely · Within months

Open Questions

  • 人工智能翻譯在什麼樣的條件下可以完全取代人類譯者?
  • 如何改進人工智能模型以更好地理解上下文和專業術語?
  • 不同語言對人工智能翻譯準確性的影響程度如何?
  • 企業在使用機器翻譯時應該採取什麼樣的質量控制措施?

Related Topics

This article was originally published by BBC中文.

Related Stories

Is Apple’s first HomeView worth buying? Foreign media lists 4 potential disadvantages
Developing·

Is Apple’s first HomeView worth buying? Foreign media lists 4 potential disadvantages

Apple has launched the 7-inch HomeView smart display, priced at US$299. It aims to become the hub of family life, integrating entertainment, communication and smart home control. However, because the price is higher than competing products, compatibility is not fully guaranteed, Siri AI is still in the beta stage, and Taiwan is not listed in the first wave of sales regions, it has caused foreign media and consumers to question its practicality and purchase value.

自由时报
2 min read
A recorder was hidden in the OpenAI high-level conference room! The nationality of the employees involved in the case was exposed, and the FBI intervened in the investigation
Developing·

A recorder was hidden in the OpenAI high-level conference room! The nationality of the employees involved in the case was exposed, and the FBI intervened in the investigation

OpenAI employee Abdulrahman Alfozan was accused of installing a hidden recording device in the company's senior executive conference room to secretly record sensitive technical communications. The 35-year-old Saudi Arabian citizen has been arrested, appeared in court for the first time and was fired. If convicted, he faces up to 5 years in prison and a $250,000 fine.

自由时报
2 min read
AI short dramas are gaining popularity in overseas markets. Chinese industry insiders suggest differentiated development
Tech·

AI short dramas are gaining popularity in overseas markets. Chinese industry insiders suggest differentiated development

AI short dramas are known as "domestic Korean dramas" and "domestic American dramas" because of their ability to generate international characters and scenes, promoting the expansion of Chinese short dramas overseas. The industry points out that AI has simplified the traditional production process, but there are also problems such as narrative routines, and differentiated development needs to be achieved through industrial ecological improvement and policy guidance. The creator emphasized that although AI has improved efficiency, the upper limit of artistic creation still relies on human beings' unique life experience and imagination.

中国新闻网
2 min read
Google's most powerful Gemini 4 Argon is now available to Pro users and above
Developing·

Google's most powerful Gemini 4 Argon is now available to Pro users and above

Google DeepMind officially expanded its flagship AI model Gemini 4 Argon on October 13, making it available to global developers, enterprises, and Google AI Pro and Ultra subscribers. A limited-time API discount is also launched simultaneously, charging USD 2 per million input tokens and USD 10 for output, half of the official price starting in December. The model supports a unified context window of up to 1,048,576 tokens and is designed for long-term reasoning and multi-step processing. Enterprises can use it through Google Cloud and Gemini for business, and developers can connect through Google AI Studio or Gemini API. The model identification code is "gemini-4-argon". Google has set the API discount until November 30. During this period, when the implicit cache conditions are met and the cache is successfully hit, you can enjoy a 95% discount by entering the token, down to US$0.1 per million. Starting from December 1, the standard price will be adjusted to $4 for input and $20 for output, which is the same as the public pricing of Anthropic Claude Opus 5.5. The Intelligence Index of Artificial Analysis, a third-party organization, shows that Argon scored 53 points under the High reasoning setting, which is lower than the 58 points of Claude Opus 5.5; however, in the Vals Index published by Google itself, Argon scored 68.9% in professional work tests such as finance, law, program development, and taxation, which is higher than its competitors. In the program development test DeepSWE v1.1, Argon leads with 77.9%, but the terminal operation and actual software engineering problem test results are better than Argon by Claude Opus 5.5 or GPT-6 Astra. Google also announced internal application results. After the engineering team implemented AI agent development through Argon, the number of code submissions increased by more than 35%, and the proportion of code submissions that required rollback was reduced by nearly 30%. The Chrome team used Argon to handle rearchitecting efforts across hundreds of files. In addition, Google is also preparing to provide 10,000 qualified scientists with six months of free quota for the highest usage level of Google Antigravity, and expand the recruitment scope of the Fairwind Program that was originally limited to security defense partners.

自由时报
3 min read
Anthropic launches new protective shield model distillation and calls out China’s “Wisdom Spectrum” for plagiarism
Developing·

Anthropic launches new protective shield model distillation and calls out China’s “Wisdom Spectrum” for plagiarism

Anthropic, an American AI company, announced new protective measures to significantly reduce the risk of competitors using "model distillation" to copy Claude's capabilities by restricting model inference data acquisition. The company also named China Zhipu AI as having tried to develop its own model in this way and has currently blocked tens of thousands of related accounts.

自由时报
2 min read
More on this topic人工智能翻譯