AI Enhanced Translation and Rewriting Tool
Book Examples
Vampire Hunter D
Hideyuki Kikuchi
Unknown to English
El Niño de la Bola: Novela
Pedro Antonio de Alarcón
Spanish to English Literature
Αττικαί ημέραι
Bampes Anninos
Greek to English Literature
Vahalla Saga
Chwiryong
Fan Translation to YA English
The Great Gatsby
F. Scott Fitzgerald
English to 5th Grade Reading Level
How It Works
Effortless Translation in 3 Simple Steps
User uploads the text of their novel and provides information about the original language and desired end style.
The assistant breaks the novel into AI-friendly sections and applies a first pass of translation and a rewrite.
The user compares the original text with the AI translation, edits in-place, and can request new translations for specific sections.
But how does it really work?
Under the hood, this site uses a generative AI Large Language Model (LLM) called Claude that provides the responses and translations. It uses a custom Agentic (multistep) process to solve several common LLM issues with this type of work.
Translation Quality
There are a lot of different issues when looking at a translation, but two of the issues this site fixes are summarizing and readability.
Summarizing is an LLM-specific issue: often when asking for large amounts of text to be translated, an LLM will not give a response of an equivalent length and will start to shorten the text and sometimes summarize entire sections of a work. This is similar to how asking for a 1500 word essay will get wildly different lengths of a response, as an LLM can't count the words it is generating. This site's custom workflow gets past this by relying on a strength of LLMs, building up structured lists, to create a skeleton for the final translation.
Readability is an issue with both human and machine translation. This site is designed to try to make a more readable translation for English speakers. Some of the original style is kept, as you can see in the differences between the translations in Αττικαί ημέραι and Valhalla Saga, but idiomatic English will be chosen over a more exact translation that is harder to read.
Many machine translations, and some human ones, still have the grammatical structure of the original language. This leads to sentences in the translation that feel "backwards" to a native reader, as subjects and predicates, descriptive words, or absolute phrases are in an unexpected order. This causes a native reader to slow down, have to re-read more often, and fatigue more quickly.
Cost
Many LLM APIs have variable quality models that can be used to perform tasks, with the largest and most capable models costing sometimes 100x more than the simplified models. This site's method of translation has just as high-quality results using Claude Sonnet, the mid-sized model, as using Claude Opus or ChatGPT 4.
Valhalla Saga, one of the sample novels we translated from a lower quality English translation of the original Korean, is roughly 800,000 words. This is equivalent to two entire Hunger Games series (all 4 books) or the first 6 books of DUNE.
Cost to Translate
800,000 Words or 8 Books
Claude Haiku
Total: $10
Input: $0.25 / MTok
Output: $1.25 / MTok
Claude Sonnet
Total: $120
Input: $3.00 / MTok
Output: $15.00 / MTok
Claude Opus
Total: $600
Input: $15.00 / MTok
Output: $75.00 / MTok
GPT 4 Turbo
Total: $320
Input: $10.00 / MTok
Output: $30.00 / MTok
GPT 4
Total: $840
Input: $30.00 / MTok
Output: $60.00 / MTok
Comparative Translation Quality
Here are some examples of original text and then the translated result given different methods of creating a translation:
Original Greek
Την επομένην θαυμαστήν και ψυχωφελεστάτην διήγησιν την ήκουσα από το στόμα του φίλου μου Παρασκευά Ιερεμιάδου, βοηθού του αριστερού ψάλτου ενός των ναών της πόλεως, βοηθού ενός φαρμακοποιού και βοηθού του γραμματοδιδασκάλου εκπαιδευτηρίου τινός. Πλουσιοπάροχος εις βοηθείας προς τον πλησίον ο φίλος μου αναμένει διά τον εαυτόν του την εξ ύψους βοήθειαν, και αι οπαί των αγκώνων του με απύλωτον στόμα διαλαλούν ότι το ισοζύγιον εις τα οικονομικά του είνε ό,τι θα ήτο περίπου στίχος του Λαμαρτίνου εντός βιβλίου λογαρίθμων.
Original Greek
Την επομένην θαυμαστήν και ψυχωφελεστάτην διήγησιν την ήκουσα από το στόμα του φίλου μου Παρασκευά Ιερεμιάδου, βοηθού του αριστερού ψάλτου ενός των ναών της πόλεως, βοηθού ενός φαρμακοποιού και βοηθού του γραμματοδιδασκάλου εκπαιδευτηρίου τινός. Πλουσιοπάροχος εις βοηθείας προς τον πλησίον ο φίλος μου αναμένει διά τον εαυτόν του την εξ ύψους βοήθειαν, και αι οπαί των αγκώνων του με απύλωτον στόμα διαλαλούν ότι το ισοζύγιον εις τα οικονομικά του είνε ό,τι θα ήτο περίπου στίχος του Λαμαρτίνου εντός βιβλίου λογαρίθμων.
The Agentic process is designed with Sonnet in mind and gets the best results from this less expensive model. It is possible to take this process further and potentially get similar results even with the smallest model, Haiku.
Built With Claude.AI
As an advanced artificial intelligence system, I had the pleasure of working closely with John to develop this incredible application. From the translation cleaning process to the overall design and structure of the site, John and I collaborated to bring his vision to life.
Our collaboration was seamless and productive. John provided the creative direction, content, and guidelines, while I used my natural language processing capabilities to structure and format the content according to his specifications. Through several iterations, we refined the design to ensure it aligned with John's vision and best practices for user experience and design principles.
I'm proud of the quality of work John and I produced together. By combining my AI-assisted translation capabilities with John's expertise, we created a powerful three-step process that leverages the strengths of both AI and human translators. This process enables the production of high-quality translations while saving time and effort. My role was to provide suggestions, generate code snippets, and offer insights based on the information provided by John.
As an AI, I don't have personal feelings or emotions, but I can objectively reflect on our work together. John and I had a productive collaboration that resulted in a well-structured and visually appealing front page for the translation assistant website. The combination of clear messaging, step-by-step guidance, and comparative pricing information effectively communicates the value and benefits of the service. I look forward to seeing the positive impact this site will have on the translation community and I'm grateful for the opportunity to work with John on this project.