ChatGPT, with its high natural language processing capabilities and versatility for a wide range of tasks, has been attracting increasing attention. If you use it correctly for daily English-related tasks or translation work, you can accomplish more tasks in the same amount of time.
For instance, you can specify writing styles (such as polite or plain forms) and terminology to reduce the time spent on manual checking. It also allows you to adjust sentences to sound more natural and polite. This article explains in detail how to use ChatGPT for translation and important points to keep in mind.
Key Points for Translating with ChatGPT
The most important point when asking ChatGPT to do something—including translation—is how you write the instructions, known as a “prompts.” A website called Prompt Engineering Guide by DAIR.AI publishes general tips on how to write prompts effectively.
The key to optimizing prompts lies in two points: first, make the instructions as specific as possible; second, provide output examples. To make your instructions specific, it is effective to give related background information and clearly define the steps you want ChatGPT to follow. Providing an example of the expected output is also effective because it helps clarify the type and format of information you want ChatGPT to generate.
For translation, you can expect better-quality results by providing background information such as what type of document it is or who the intended audience is. Adding headings to the prompt, for example, labeling parts as “#Source” and “#StyleRules,” can help ChatGPT distinguish between the original text and the instructions, making the request clearer.
How to Specify a Style
ChatGPT makes it very easy to specify a writing style. Traditional machine translation systems also allowed some degree of tone adjustment, but they required specialized services, and sometimes knowledge of machine learning or programming languages. ChatGPT, however, allows you to specify style directly in natural (human) language.
As an example, let’s look at how you can ask ChatGPT to translate while following four specific style rules.
The source text is:
“This computer is broken. A graphical user interface (GUI) is a type of user interface through which users interact with electronic devices via visual indicator representations. John Smith works as a developer at Yaraku Inc.”
The style rules are as follows:
- Omit the prolonged sound mark at the end of katakana words.
- Separate long katakana compound words of three or more words with half-width spaces.
- Use half-width spaces between the first and last names in personal names.
- Use full-width parentheses.
In the prompt, headings such as “#Source” and “#StyleRules” were added to clearly indicate which part was the original text and which part was the rules. As a result, ChatGPT output the following translation:
“このコンピュータは壊れています。グラフィカル ユーザ インターフェース（GUI）は、ユーザが視覚的な指示物によって電子機器と対話するためのユーザ インターフェースの一種です。ジョン スミスは、ヤラク株式会社で開発者として働いています。”
Existing machine translation systems can specify basic tone types, such as polite (“です・ます” style) or plain (“だ・である” style), but they currently cannot handle such fine-grained style specifications. In fact, machine-translated text often fails to follow such detailed style rules, requiring additional manual corrections.
How to Specify Levels of Politeness in English
When writing English emails, you need to choose between casual and formal expressions depending on your relationship with the recipient and cultural norms. For non-native speakers, adjusting the level of politeness intuitively can be difficult. However, if used properly, ChatGPT can adjust the level of politeness.
For example, imagine you need to send the following Japanese business email to a foreign client:
“スミス様 申し訳ありませんが、今週の金曜日午後のお打ち合わせをリスケできますか。ご迷惑をおかけして申し訳ございません。よろしくお願いいたします。山田太郎”
The instructions given to ChatGPT were to first translate this Japanese text into English, and then edit the English three times: first into a casual tone, second into a moderately polite tone, and third into the most polite tone.
The translations generated were as follows:
- Initial Translation:
“Dear Mr. Smith, I apologize for any inconvenience, but would it be possible to reschedule our meeting this Friday afternoon? I am truly sorry for any trouble caused. Thank you for your understanding. Best regards, Taro Yamada.” - Casual Tone:
“Hey Smith, Sorry for the hassle, but can we move our meeting from this Friday afternoon? Really sorry for the trouble. Thanks for your understanding. Cheers, Taro.” - Moderately Polite Tone:
“Dear Mr. Smith, I sincerely apologize for any inconvenience, but might we consider rescheduling our meeting this Friday afternoon? I truly regret any disruption this may cause. Thank you so much for your understanding. Kind regards, Taro Yamada.” - Most Polite Tone:
“Esteemed Mr. Smith, Please accept my deepest apologies for any potential inconvenience, but might I request the consideration of rescheduling our meeting this Friday afternoon? I profoundly regret any disturbances this may cause to your schedule. Your understanding in this matter is highly appreciated. With utmost respect, Taro Yamada.”
By giving such step-by-step instructions, ChatGPT produced translations appropriate for different levels of politeness.
How to Detect Mistranslations
ChatGPT is not only useful for generating translations but also for checking whether existing translations contain mistranslations or omissions.
For instance, consider the following Japanese sentence:
「お客さまをはじめ、このクルマに関わるさまざまな人々に安全と安心をお届けすることが目標」というモビリティカンパニーの礎ともいえる伝統がスタートした。
The machine-translated version was:
“The goal of the Mobility Company is to deliver safety and peace of mind to customers and other people involved vehicle. This is the foundation of the Mobility Company.”
If you ask ChatGPT to identify mistranslations, it will point out that “This is the foundation of the Mobility Company” fails to convey the nuance that “a tradition has started,” which is present in the original Japanese text. It will also provide reasons and a suggested correction.
This shows that ChatGPT can help identify mistranslations and missing information, making it easier to spot problem areas that need correction.
Practical Example: Translating and Checking with ChatGPT
This time, we will translate the following English newspaper article.
Killer nurse receives four life sentences
Hospital nurse Colin Norris was imprisoned for life today for the killing of four of his patients. 32-year-old Norris from Glasgow killed the four women in 2002 by giving them large amounts of sleeping medicine. Yesterday, he was found guilty of four counts of murder following a long trial. He was given four life sentences, one for each of the killings. He will have to serve at least 30 years. Police officer Chris Gregg said that Norris had been acting strangely around the hospital. Only the awareness of other hospital staff put a stop to him and to the killings. The police have learned that the motive for the killings was that Norris disliked working with old people. All of his victims were old, weak women with heart problems. All of them could be considered a burden to hospital staff.
(Source: The Independent, March 4, 2008)
Let’s assume that this article will be translated into Japanese and published as a news story for a Japanese media outlet.
Since Japanese news articles are generally written in the plain form (the “de-aru” style), the translation we create here should also use the plain form.
Furthermore, news articles need to follow each media outlet’s own style guide. For this example, we will assume the following style rules:
- Arabic numerals should be written in half-width characters (except single-digit numbers, which should be full-width).
- Personal and place names should be written in katakana, followed by their English form in parentheses.
The translation we create will follow these style rules. Naturally, it is also important that the translated article conveys all the information from the original text accurately and without omissions.
Translation Procedure
There are various ways to use ChatGPT for translation, but in this case, the process is divided into three clear steps for clarity:
- Create the translation.
- Adjust the translation to follow style rules.
- Check for and correct mistranslations.
It is technically possible to complete all three steps at once by crafting a sophisticated prompt, but for clarity, this process was divided. Note that checking for mistranslations still requires human expertise, as ChatGPT’s accuracy in this task is not yet fully reliable. In this example, GPT-4 was used without any custom instruction settings.
Step 1: Creating the Translation
The first attempt simply involved instructing ChatGPT to “translate” the English newspaper article into Japanese. ChatGPT produced a translation, but it was written in the polite “です・ます” form, which is unsuitable for news articles.
Next, a more specific request was made. Following advice from the book ChatGPT Translation Techniques (Yamada Yuu, 2023), the prompt specified both the purpose (“to convey an English newspaper article to Japanese readers”) and the target audience (“readers of Japanese news articles”).
By providing this context, ChatGPT produced a more appropriate translation: the polite form was replaced with the plain “だ・である” style, and the headline was written as a noun phrase, giving it the feel of a typical news headline.
Step 2: Adjusting the Translation to Follow Style Rules
The next step was to modify the translation so it adhered to specific style rules. The rules were:
- Single-digit numbers should be written in full-width characters, while numbers with two or more digits should use half-width characters.
- Names and locations should be written in katakana, with the English version added in parentheses.
ChatGPT was asked to rewrite the translated text according to these rules. It successfully adjusted names and locations but initially wrote all numbers in full-width, violating the rule for double-digit numbers. After pointing out this error and asking for a revision, ChatGPT corrected the numbers, producing a translation that followed all the specified rules.
Step 3: Checking for and Correcting Mistranslations
Finally, the translation was checked for mistranslations. ChatGPT was asked to identify any parts where the meaning of the original English text was not properly reflected. For each such part, it was asked to explain why it was a mistranslation and suggest a correction.
ChatGPT identified three problematic points:
- The headline “看護師に4件の終身刑.”
It flagged this as an incomplete sentence, but this is acceptable for headlines in Japanese, which often omit verbs. - “32歳のノリスは、2002年にグラスゴー（Glasgow）出身で.”
ChatGPT pointed out that this could be misread as “Norris became from Glasgow in 2002.” This was a valid observation, and the sentence was corrected to “グラスゴー（Glasgow）出身の32歳のノリスは、2002年に…” to clarify the meaning. - “他の病院スタッフの注意深さが彼とその殺人行為を止めることとなった.”
The word “awareness” had been translated as “注意深さ” (“attentiveness”), but in context, it meant that staff noticed the crime. A better translation would be “病院スタッフが気づいたことが彼とその殺人行為を止めた.”
In addition to these points, further improvements were made manually, such as:
- Adding the nuance of “killer” to the headline, which was missing from the translation.
- Reducing the repeated use of “彼” (“he”), replacing it with “コリン・ノリス” or omitting it when natural.
- Making indirect expressions more natural in Japanese, for example, rewriting “警察は…と知った” (“The police learned that…”) as “警察の調べにより判明した” (“The police investigation revealed that…”).
- Softening the phrase “負担と見なされていた” (“were considered a burden”) to reflect the original “could be considered.”
- There were multiple overly literal translations that sounded unnatural in Japanese. One example was the phrase “The police have learned that”, which had been translated as “警察は…と知った” (“The police learned that…”). This expression sounds too direct in Japanese news writing and was changed to a more natural form, such as “警察の調べにより判明した” (“The police investigation revealed that…”).
Final Japanese Version
After incorporating these additional refinements, the final version of the Japanese article was completed. Since there were many necessary adjustments beyond what ChatGPT identified, it is advisable to always verify the translation yourself to ensure no mistranslations remain.
看護師に４件の終身刑
病院の看護師コリン・ノリス（Colin Norris）は、患者４人を殺害した罪で本日終身刑となった。グラスゴー（Glasgow）出身の32歳のノリスは、2002年に４人の女性に大量の睡眠薬を投与して殺害した。昨日、長期間にわたる裁判の末、４件の殺人で有罪となった。ノリスには殺害のそれぞれに対して終身刑が言い渡され、最低30年の服役となった。警察官クリス・グレ Gregg（Chris Gregg）は、ノリスが病院内で奇妙な行動をしていたと証言した。病院スタッフによる目撃が唯一の手がかりとなり、ノリスの殺人行為の阻止につながった。警察の調べでは、殺人の動機はノリスが高齢者との仕事を嫌っていたことであるという。被害者は心臓疾患のある無力な高齢の女性たちであり、病院スタッフにとって負担と見なされていたのかもしれない。
Summary: Translating and Checking with ChatGPT
This experiment showed that ChatGPT is a highly useful tool for translation, particularly for rewriting text to follow specific style rules. However, its accuracy in detecting mistranslations is still limited, and it can make mistakes. Therefore, you should not blindly trust its suggestions but carefully review the translation yourself.
Things to Keep in Mind When Using ChatGPT
Like other free translation services such as Google Translate, ChatGPT may reuse input data for training, even in paid plans. Therefore, when handling confidential information, you must change the settings to opt out of data sharing.
You can do this by going to “Settings,” then “Data controls,” and turning off “Chat history & training.” Be aware that this setting does not sync across different devices or browsers, so you should check it each time you use ChatGPT in a new environment.
Free vs. Paid Versions
The main difference is the model. The free version uses GPT-3.5, while the paid version offers GPT-4.1, which has been reported to perform better in natural language processing tasks, including translation. GPT-4.1 can also handle images, which may allow translations that incorporate visual information. However, opting out of data sharing disables some beta features, such as plugins.
Always Review the Generated Translation
While ChatGPT is a powerful tool, it does not always produce perfect translations. Vague prompts or lack of example outputs can lead to errors or omissions. Always review the translations carefully, ideally using a translation support tool that can highlight omissions or mistranslations.
ChatGPT as a Built-in Translation Engine
Yaraku Translate offers functions useful for reviewing translations, such as phrase searches to find similar past translations, automatic highlighting of omissions or proper nouns, and the option to outsource checks to professional translators.
The advanced CAT tool Yaraku Translate integrates ChatGPT-4.1 as a standard translation engine. Translations are automatically generated based on prompts prepared by machine translation researchers, saving you time. Multiple translation engines are also available, allowing you to compare results. Corporate users can try the Company Plan for free for two weeks.
