Table of contents
- From Machine Translation to AI Translation
- The Difficulty of Literary Translation
- Using AI for Draft Translations, but Publishing at the Draft Stage
From Machine Translation to AI Translation
I previously wrote about an idea using NotebookLM, suggesting that “a project where AI creates draft translations of Aozora Bunko, which are then sequentially translated by human translators, might deserve budget priority over Cool Japan.” Just to be clear, I am not involved with Aozora Bunko itself. This is a discussion about whether the archives of Aozora Bunko can be utilized by a third party.
Before I started writing novels, I worked in the translation industry, where I possessed know-how for systematically and efficiently handling large volumes of corporate translation and improving translation quality. I have also contributed to magazines for translators and English education magazines for many years. Based on that experience, I believe that AI draft translation of literature is meaningful.
Automatic translation used to be called “machine translation.” It was reasonably useful with appropriate preparation, environment, and post-editing, but without those, it could only produce insufficient translation quality. This was improved by a method called statistical translation, and then neural machine translation appeared, eventually becoming more sophisticated as the current automatic translation. There are still terrible automatic translation web pages and translations (localization). However, in the end, whether the quality of automatic translation can be improved depends on how it is used.
The Difficulty of Literary Translation
Literary translation is a very difficult task. It requires deep scholarship, including an understanding of the author themselves, as well as the work and its historical background. However, even for literary experts, self-taught translations without translation training are often not polished and can be difficult to read. And with the exception of some bestsellers, literary translation is rarely economically viable. As a result, valuable Japanese literature remains buried without being known to the world.
Until now, technical limitations of AI sometimes made it impossible to translate long texts at once. Translating in bits and pieces leads to a lack of contextual information, resulting in incomplete translations. But now, long-text translation has become possible. By using RAG like NotebookLM, anyone can create fairly high-quality draft translations almost for free. Translation quality can be further improved by generating glossaries and performing literary analysis. A simple example is below.
Related: NotebookLM for literary translation
Mistranslations will likely increase in cases of delicate styles like Izumi Kyoka or somewhat special phrasing and vocabulary like Miyazawa Kenji, and sufficient quality may not yet be achieved in poetry. However, I believe it is meaningful to make Aozora Bunko readable with a small budget so that people around the world can learn about Japanese culture through Japanese literature.
Using AI for Draft Translations, but Publishing at the Draft Stage
Using AI for draft translations, but publishing them at the draft stage—this is the point of this proposal. Ultimately, it is expected that human translations by literary translator volunteers will be included. The immediate goal is to translate Aozora Bunko into English, and then into multiple languages.
It is also possible to build a translation system to improve draft translations. This would be a so-called translation memory, decomposed sentence by sentence and editable in a bilingual state. I tried making the base app for this using Vibe Coding.
I want to start with an experiment to see what can be done. In the future, it would be great to expand to all languages with a service that could be called Wiki Literature.
Please let me know if any interested universities, companies, or organizations would like to get involved.