Chat better, translate better

Some reflections on AI translation

This electronic newsletter specializing in English translations of my works started last September, and has since then serialized chapters from Hong Kong Type, Unnatural Recollections and the Bodhisattva stories. The first and the last have already been completed, and publication in NFT book form is forthcoming.

As I have said in the opening post, “At lease the raft, if not the shore”, the use of AI translation is a necessary means for the autonomy and empowerment of the author in face of extremely limited resources and opportunities to be translated and published in the ordinary way. Even though the limitations of AI translation are obvious, and self-publishing is far from being an accepted or even respected way of dissemination, it is always better to have reasonably accurate and readable translations than to have none.

Let me reiterate: translation is an art in itself and top human translators are always invaluable and irreplaceable. Yet such translators are often out of reach for most writers of minor literatures, and if they want to have the fair chance of being read outside their own native language, technology is their only option. AI should neither be deified nor demonized. Used correctly, it is a useful tool, and extremely so.

My experiment with AI translation is far from comprehensive. I have been using only GPT-4 of OpenAI and have no knowledge of how other models perform. Since first trying my hand at AI translation a year ago, GPT itself has undergone great transformation, and a more powerful version GPT-4o has just been launched. However, even with the same model, when it is accessed through different user interfaces, it will yield results of different qualities.

At the very beginning, I used the ChatGPT online chat box for translation, which requires the user to paste texts in chunks no longer than a few hundred words, or else the system will cease functioning. The consistency between requests is far from satisfactory, with noticeable shifts in style and diction, and huge amount of editing work is needed to put the fragments together into a unified whole. It is rather clumsy and time consuming. I call this the manual input method.

Then I tried using OpenAI’s API, together with an open-source software called Bilingual Book Maker. This is a powerful tool, with which a whole book can be translated in a matter of a few hours, at the cost of a few USD. I used this fully automated method to translate the first draft of Hong Kong Type and Unnatural Recollections. While being extremely efficient, the quality is unstable. It is virtually impossible to give detailed instructions and the generation goes somewhat randomly, mixing brilliance with mediocrity, if not downright mistakes. Its major problem is decontextualization, which results in a lack of stylistic consistency and unity, similar to what happens in manual pasting and prompting. The editing work is heavy and some fundamental flaws in linguistic style and tone can hardly be rectified unless redoing it all over. Since I haven’t applied parameters like “temperature” to the prompt, the language sounds formal and a little unnatural for a literary work.

To be fair, when used in Chinese to English translation, GPT-4 is accurate and faithful to the original, but at the same time very literal. It is enough to convey what the work is about to a foreign reader, but the artistry of the language is lost. So, it seems more suited to translating factual and practical works than literary works, and it requires considerable human intervention to remedy the drawbacks. The editing work on Hong Kong Type was quite formidable and the result far from perfect. But I believe that it is a question of the working method, rather than some inherent and insurmountable weaknesses of AI translation itself. How to use AI remains something to explore and learn.

Recently, I setup a personal GPT for the purpose of translation. It is different from just pasting text into the chat box. The translation helper is tailor-made for this specific purpose. As it says, “This GPT assists users in translating their works accurately and efficiently. It analyzes the style of the original text and considers cultural nuances, idiomatic expressions, and context-specific terminology. It provides translations and explanations, ensuring clarity and precision.”

This time I tried a totally different procedure: First upload the source file to be translated, so far only one short story a time, then ask ChatGPT to analyze the language and style of the piece, before proceeding to carry out the translation task. Afterwards, I will evaluate the result and ask further questions to direct it to issues that need further refinement. I discover that ChatGPT displays considerable power in the interpretation of meanings and references, or at least a semblance of it, which is the prerequisite for good translations. With such guiding principles, I find it more focused on its task and gives better results. It looks like it understands what it is doing.

I have used this method with nine short stories from Unnatural Recollections and six pieces from a new series SF’s Strange Daily Life. Some of its analyses are amazing and the translated texts reflect a better grasp of the meaning, both literal and metaphorical. Its ability to discern humor and irony, and to differentiate between emotional states, is impressive. It even claims to have made a choice because it is more poetic. Actually, it can give (or make up) reasons for every choice it makes in a rendition.

My temporary conclusion is that, to boost AI’s performance, the user needs to be more involved in the process and actively interact with it. If left alone, it will run offtrack. All it needs is attention. The more input and the more precise the instructions you give, the more directed it is to a certain task. But that means one has to do the translation by shorter units, in chapters or sections, so as to closely monitor the output. Producing a book length translation in one go is not a recommendable practice. Automation has to be paired with human judgment for the best results.

These are just some personal reflections on the use of AI translation, based on my limited scope of experiment. Rather than gaining any insight in anything, the only thing I have learned is that there are immense possibilities of AI usages which we haven’t yet explored and discovered. We should learn to tap this potential for the benefit of expanding the capacities of literary creations.

For readers who are interested in how AI translation works, at least in one of its many ways, I will post these recent experiments in full, including not just the translated text but also the analyses, explanations and conclusions ChatGPT makes, in response to my prompting and querying. The Chinese original is also attached at the end for reference. The performance of these tasks is remarkable - over 99% completeness, which means in most cases not even a single word needs to be altered.

*This article is directly written in English, not translated or polished up by AI.