Free online consultation available. Please feel free to contact us.

Self-translate prompting
"Think in English and then answer in Japanese."

  • Glossary

Self-translate prompting /"Think in English, then answer in Japanese"

This glossary explains various keywords that will help you understand the mindset necessary for data utilization and successful DX.
This time, we will introduce "Self-translate prompting," a technique for utilizing conversational AI using large-scale language models such as ChatGPT.

What is Self-Translate Prompting?

Self-translate prompting is a prompt creation technique that allows you to specify the language to be used for input (prompts) and output, as well as the language ChatGPT uses in its thought process, when using conversational AI such as ChatGPT, to obtain more desirable output.
A well-known example is adding the phrase "think in English, then answer in Japanese." In this case, users can get improved output while still using the language as usual, using only Japanese.

What kind of technique is it? (Example of use)

This technique was originally popular right after ChatGPT was released, when performance was not as good as it is now. Since then, performance has improved, so it is no longer necessary to use it, but I think it is now being used in different ways than before.

It is recommended to use it in English (but it is troublesome)

When ChatGPT first became popular, I was often told to "try using it in English." I tried it out but didn't get a satisfactory answer, so I would say, "ChatGPT wasn't as good as I expected," and I was often told that it was because I wasn't using it in English.

Please tell us about the importance of EAI tools in utilizing data.

EAI tools are an important technology that enables data integration between different application systems within a company, and are important tools for integrating different data to support accurate decision-making.

For example, suppose the above wasn't enough.

Please tell me about the importance of EAI tools in data utilization.

EAI tools are important in data utilization because they enable data integration between different enterprise application systems, promote information sharing, maintain data integrity, and support efficient decision-making.

In this example (and at the time of writing), there doesn't seem to be much difference. However, at the time, it was said that "using it in English makes you smarter than a university student" and "using it in Japanese makes you smarter than a high school student."

It's hard to argue against the idea that we should use English (especially from people who can speak English). However, typing in English is more mentally tiring than typing in Japanese, and even if you use machine translation (as we did above), it still takes time to translate. ChatGPT is characterized by its ease of use, so it's unfortunate that it has to be more troublesome.

Introducing techniques that can be used in Japanese

What was helpful in this situation was the "method of using Japanese as the English equivalent." It is used almost the same way as normal Japanese, but you add the phrase "think in English and answer in Japanese" to create a prompt.

Please tell us about the importance of EAI tools in utilizing data.
Think in English and answer in Japanese.

Enterprise Application Integration (EAI) tools are important tools for enabling data integration between different applications, and are used to collect data from different data sources, centralize it, and provide a consistent view.

When it comes to utilizing data, EAI tools are important for the following reasons:

  • Centralized Data Management: EAI tools allow you to centralize data from multiple data sources, eliminating the need to access scattered data and improving productivity.
  • Improved data quality: EAI tools can improve data quality. Centralized data management can resolve data quality issues, which in turn leads to accurate data being used for better decision-making.
  • Real-time data processing: EAI tools can enable real-time data processing, which allows for faster data processing, such as analysis and report generation, saving time.
  • Data Security: EAI tools can ensure data security. They enable security controls such as access control and data encryption.

As mentioned above, EAI tools are extremely important in utilizing data and are one of the essential tools for efficient business operations.

Compared to using English yourself, you simply add phrases, so it is a technique that allows you to bring out the same abilities you would use in English while maintaining the convenience of using it comfortably in Japanese.

The answer may or may not differ much. The example output above is a copy of the answer I got after trying several times.

What does Self-translate prompting do?

Why does ChatGPT give better answers in English? ChatGPT is technically realized using a large-scale language model, and it is not clear how it actually works internally.

​⇒Large Language Model (LLM) | Glossary

However, of the vast amounts of text data used to train large-scale language models, the largest amount of data is written in English, and the majority of training is done in English, so it is thought that this gives the models the highest performance. At the very least, it is unlikely that there would be more Japanese text data than English and that there would be a sufficient amount of training, especially for large-scale language models created overseas. If there are differences in results, it should be assumed that this is due to the large amount of training done on English data.

Specify the language you think in and the language you respond in

If you think about it, it's quite strange that a large-scale language model would behave in this way. By simply adding the phrase "Think in English and answer in Japanese," it seems that the model somehow understood that we wanted the model to respond in Japanese, just as it would if it were in English, and behaved in this way.

It has not been trained to behave in this way, as it has not been individually trained to have this ability simply by feeding it a large amount of text data.

Furthermore, rather than just using this as a technique that can improve performance simply by adding a "magic phrase," if we think of it as a technique for controlling the behavior of large-scale language models, by allowing us to specify the language used for communication with users and the language used for thinking, it may be possible to use it for a wider range of purposes.

For example, even if you're using English, there are likely to be questions for which the quality of the answer would be better if asked in Japanese. For example, if you're talking about anime, manga, or Japanese culture itself, the text data used for training will likely be mostly Japanese. In that case, adding a phrase like "Think in Japanese and answer in English" may improve performance.

This technique should not be limited to English or Japanese. It could be used in a variety of ways to generate desired answers, such as adding a phrase like, "If you want to ask about a French philosopher, think in French and answer in Japanese."

You, the user, control how machine learning works

In conventional machine learning, training data is prepared in advance and the trained model is then basically used as is. In any case, anything beyond using a prepared trained model as is was an area that only highly skilled engineers could get involved in.

However, with large-scale language models, users have the ability to change the language used for thinking simply by adding phrases in natural language. Adding phrases is not just a technique for improving performance, but also a way to interfere with the AI's behavior. If you think of it as a revolutionary thing, it allows you to do advanced things that you previously couldn't do yourself.

It depends on the situation in which you use it to improve the output, but if you understand this technique, you can think of it as being able to change the behavior of ChatGPT on a qualitatively different level.

Keywords related to Generative AI/ChatGPT (for further understanding)

Glossary Column List

Alphanumeric characters and symbols

A row

Ka row

Sa row

Ta row

Na row

Ha row

Ma row

Ya row

Ra row

Wa row

»Data Utilization Column List

Recommended Content

Related Content

Return to column list