Prompting for Patent Translation

Effective AI translation prompting, whether for translation generation, translation interrogation, or translation quality evaluation, requires an understanding of how large language models work. Media reports about models going rogue and defying their owners and talk of AI consciousness make it easy to think of LLMs as people, who understand what needs to be done and use judgment and creativity to achieve it. But descriptions of AI as thinking, planning, deciding, trying, wanting, etc., should be understood as analogies, similar to how we say that water “seeks” the lowest level, or that plants produce nectar because they “want” to attract insects.

A large language model is a mathematical function that returns a value, just as a function f, which takes x as its input, returns the value y. It's a complex function with billions or even trillions of parameters, and it runs iteratively, taking its own output as further input, and deliberately mixing in some randomness, but it's still a function. As such, the output depends on the input. Recall from high school math that in f(x) = y, the value of y depends entirely on the value of x. In LLMs, x is the prompt. In other words, the output of an LLM depends on the prompt (and on the amount of deliberate randomness that has been set).

Because an LLM is a mathematical function, and not in any way an artificial human brain, it cannot remember, and it cannot learn. Every time that you hit Enter in a chat interface, you are sending a brand-new prompt to a model that doesn't know anything about you or what you want. It will generally seem like the model remembers previous portions of the conversation because every time you enter something new into a web interface chat session, the interface also sends the entire conversation up to that point. In other words, when you type your second prompt and hit Enter, the model receives all of your first prompt, all of its own response, and all of your second prompt; these three things together become the input x.

With that in mind, we arrive at our first best practice, which is prompt hygiene.

Prompt hygiene

Consider the case (or try it yourself) in which you ask an AI:

Can you help me find a patent published prior to 2024 which includes an embodiment for suppressing PCB delamination due to thermal shock?

and, later in the same conversation, you ask it to:

Translate この形態では、剥離しやすくなる into English.

Gemini and ChatGPT, for example, return:

In this embodiment, delamination becomes more likely.

Now consider what happens if instead you first ask:

Can you help me find a patent published prior to 2024 which includes an embodiment in which a silicone spray is used for mold release?

and then later paste in exactly the same translation request (Translate この形態では、剥離しやすくなる into English). Now those models respond:

In this embodiment, it becomes easier to release.

In one sense, this is great. The AI shapes its translation to match the field that you are working in. But if you are working on more than one matter or there is more than one aspect to the case you are researching, it's easy to unintentionally influence the translation. This is known as context poisoning. Prompt hygiene means being deliberate about the value of x that is passed to the model, in the knowledge that the value of y will be determined by it.

The flip side of context poisoning is context starvation. If you open a fresh session and, without any prior prompting, ask it to translate the same sentence, この形態では、剥離しやすくなる, into English, it returns:

In this form, it becomes easier to peel off.

The context-starved model, with nothing to go on, falls back on the most statistically common meanings for 形態 (form, rather than embodiment) and 剥離 (peeling off, rather than delamination or release).

In other words, the same models gave three different translations for exactly the same text depending on the prompt context. In each case the model returned the value that the input determined. Whether or not these constitute translation errors depends on the context.

Hidden prompts

A new conversation in a chat interface seems to start from zero. But recent consumer interfaces for models such as ChatGPT, Gemini, and Claude all offer some form of persistent memory, meaning that notes about you, your work, and your past conversations are stored by the service and inserted into the input, alongside every new prompt.

If the service has stored the facts that you are a litigator and that your firm represents a manufacturer of printed circuit boards, this context may help (or poison) your translation prompt. In some setups, custom instructions and system prompts are also sent to the model and can influence the translation that it returns.

Short of using an API-based front end such as TypingMind to control exactly how you interact with the LLM, it is generally possible to disable the memory on frontier model interfaces, such as by selecting an incognito mode, but the specific settings depend on the service. If disabling memory is not convenient, bear in mind that the translation that the system returns may not be neutral, even with a fresh prompt.

Settings

It should go without saying that you do not want to try to translate a Japanese patent with a free model on a public website. Choose a paid flagship model that the provider advertises as its most capable, avoiding ones that are touted as being inexpensive or fast. If the model you use allows you to set the temperature (the amount of deliberate randomness), choose a low temperature such as 0.1. If you have the option of setting the thinking budget within the top-end model that you use, set it to low or medium. Setting the thinking budget too high can cause the model to give too much weight to its own output and too little to the text that it is supposed to be translating.

Simple, deliberate prompting

When professional translation services make calls to large language models, they often use prompts that are thousands of words long and include complex data structures. For nonspecialist users, however, a simple structure is not only sufficient but usually superior. Complex prompts without proper structuring carry a risk of unintentional poisoning, and they can also produce excessive “thinking,” which can result in omissions and hallucinations.

Start by explicitly describing how you want the model to behave.

You are an expert patent translator translating a Japanese patent in the field of ([semiconductor fabrication], [genetic engineering], …) into English. A patent attorney who does not read Japanese will use the translation to assess what the document discloses for the purpose of ([preparation for litigation], [response to an office action], …).

Be careful not to influence the model by telling it what you want the outcome to be. If you inadvertently let it know that you want the translation to show something in particular, it will be happy to comply, regardless of the facts of the matter.

As we have seen, the correct translation depends on context. Even if you only need a single sentence or paragraph translated, provide the model with the entire document in the original language as context. Most model interfaces allow you to attach files, but it is also fine to paste the entire patent in.

I will ask you to translate ([one sentence], [two embodiments], …) from the patent publication but you must consider the entire patent, which I have provided, for context.

It's tempting to add even more reference material, the way that you might do when requesting a translation from a human translator. For example, you may have a number of related documents in English that could be helpful for getting the vocabulary right, or you might think about adding other translations that you consider to be well done, as examples. However, the motto of an LLM is that the prompter is always right. Unlike a human translator, an LLM will not set aside the context provided, even when that context turns out to be unsuitable. It is not impossible to address this with complex prompting, but the easiest way to avoid biasing AI is not to place your finger on the scale in the first place.

You can, however, give the model some guidance as to how to approach the translation. By default, LLMs are likely to return a functionally equivalent translation rather than a literal translation. This is useful if you are only interested in the gist, but it may not be faithful to the actual disclosure. You can spell out what you need in simple terms.

Keep the translation faithful to the wording used in the original document. Do not add or omit anything.

You can also provide it with a guide, such as this book chapter that sets out best practices for the literal translation of patents.

Follow the translation methodology outlined in https://patenttranslations.com/publications/handbook.html

(As some models may not be able to access URLs directly, you may wish to download the PDF and attach it as a file.)

With the context properly outlined, you can add the actual text to be translated. It is usually best to enter the entire prompt at once. Entering part of a prompt and allowing the model to reply before continuing with additional instructions may degrade your translation. It is fine to enter multiple paragraphs into the interface, using Shift+Enter to insert paragraph breaks, and to attach multiple files.

You are an expert patent translator translating a Japanese patent …
…
…
Translate ([この形態では、剥離しやすくなる], [jp2000123456.pdf, attached], …) into English.

The results will depend greatly on factors beyond your control, such as the clarity with which the application was written, the extent to which the claimed technology has been described in the literature, and the biases that the model may have acquired during training. All of the risks discussed in the first note in this series still apply. But the results of prompting a large language model in this way are generally superior to conventional machine translation.

The next note in this series will cover how to interrogate an LLM about specific matters in a translation.