ebt-deepl

ebt-deepl

The ebt-deepl repository is a Javascript library for Early Buddhist Texts (EBT) translation using DeepL.

Translation is difficult. It is difficult because definitions are fluid as they are adapted to communicate societal needs and desires. And when societal needs and desires change quickly, as they are today, the translation also has to deal with the communication gaps between living generations within a single society. For an internet world, that translation problem is global. Translating texts takes time. Translation takes years. And in those long years, a new generation will be born, a new generation that will grow up to create their own terms and catch phrases, unknowing and unwilling to learn the words for the needs and desires of older generations.

Yet in that stormy flood of change, there are invariants. The invariants that remain do so regardless of the names by with which they are painted. Praised, reviled, ignored, remembered or forgotten, those invariants stand true through the generations, through the ages.

And, if we are to understand each other, we have to start with those invariants. Indeed, reliable communication relies on shared, verifiable truths. The security of the internet itself relies on shared verifiable truths. For the internet, those truths are SSL certificates. But to understand each other meaningfully, we need to talk about meaningful truths.

What is meaningful?

Well, judging from the internet, there appears to be a lot of meaningless drama in the world. There is a lot of suffering. And if suffering is a problem, then if there was an end to suffering, perhaps that might be something good to talk about. Indeed, the internet devotes a lot of bandwidth to various purported means of ending suffering. We have pop-up ads and we have math videos. Oddly, pop-up ads claim to end suffering but rarely do, whereas math videos exclude suffering from discourse since suffering is not a computable axiom.

Interestingly, thousands of years ago, someone did talk about noble truths. In fact, that very notable person did call them Noble Truths and declared them to be self-verifiable. And thousands of years later, people are still finding those Noble Truths meaningful.

DN1:1.3.2: “It’s incredible, reverends, it’s amazing how the diverse convictions of sentient beings have been clearly comprehended by the Blessed One, who knows and sees, the perfected one, the fully awakened Buddha.

One of the most fascinating things about these Noble Truths is that they have survived and flourished with repeated translation. They have remained meaningful through generations and ages. In other words, the Four Noble Truths, might be a remarkable basis for translation and communication about meaningful topics. So although we could talk about things like painting ourselves to look young, we might instead wish to talk about ending suffering reliably and verifiably.

But wait. We need more than four truths for translation. Bummer.

Well, math found a solution to a paucity of truth. It's called the proof. Math has few axioms and many proofs that extend those axioms in useful ways that mathematicians discuss endlessly.

Can the four Noble Truths be extended in the same way? Well, the EBTs do claim so:

MN80:16.5: Let a sensible person come—neither devious nor deceitful, a person of integrity. I teach and instruct them.
MN80:16.6: Practicing as instructed they will soon know and see for themselves,

It would appear then that the Early Buddhist Texts might prove to be a consistent, reliable, and verifiable candidate for a contemporary translation basis.

And why does that matter?

Well, if we take a consistent, reliable, and verifiable basis for contemporary translation and feed it to a machine translator, we might be able to understand each other better and talk about finding meaningful solutions to meaningful things...

DeepL

DeepL is a state-of-the-art automated translator. DeepL has a glossary feature that supports the customization of vocabulary for selected purposes. And the glossary feature is remarkably well-suited for representing a large translation basis. DeepL allows for up to 5000 glossary entries, which provides a customizable basis of up to a solid B2 level

And with DeepL customized this way, we could talk together, if we wanted to, about everyday things and about meaningful things.

iti75:5.2: It’s when some person gives to everyone—whether ascetics and brahmins, paupers, vagrants, nomads, or beggars—such things as food, drink, clothing, vehicles; garlands, perfumes, and makeup; and bed, house, and lighting.

The EBT-DeepL project

The EBT-DeepL project is still in a research phase researching how to create a reliable translation basis using the EBTs. In particular, we have found it quite useful to rely on the semantic web of meaning discovered in curated EBT examples. The EBT examples guide EBT readers through modern translations of the Pali Early Buddhist Texts found on SuttaCental.

What the EBT examples reveal is the remarkable attention to consistency in the authoring of the prose EBTs. Terms are repeated throughout the EBTs and used quite consistently with remarkable nuance, in contexts both broad and deep. Searching for EBT examples reveals an amazingly rich semantic web that has assisted and guided ongoing human translation efforts over the ages. Indeed, the process of translation itself fosters additional discussions which reveal deeper connections within the EBTs. Because of this, we rely on the EBT examples to guide our efforts in building the EBT-DeepL glossaries.

Briefly, our process is simple:

• Choose an example
• Choose a short document with that example
• Compare the DeepL translation with the human EBT translation
• Update the DeepL glossaries with the guidance of human EBT translators
• Publicly provide DeepL translations with MIT License, thereby allowing human EBT translators to refine their own translations.
• Repeat with another example

Finally, we are very grateful to Bhante Sujato and the entire SuttaCentral community for their joint work in continuosly translating the Early Buddhist Texts for all to read.

Create DeepL account

• Login to your DeepL account and use the top-right menu to open your Account settings
• Choose the Account tab to show your authentication key
• Copy the authentication key

Clone and test repositiory

git clone https://github.com/sc-voice/ebt-deepl.git
npm install
mkdir local
echo YOUR_AUTHENTICATION_KEY > local/deepl.auth
npm run test

Translate a sutta

EBT-DeepL translates from two sources having consistent and extensive Pali EBT coverage. The first source by default is Bhante Sujato's EN translations. The second source by default is Ayya Sabbamitta's DE translations. DeepL translations will be provided for both translation sources. A human reference is also useful--by default the reference author will be a Bilara author having existing consistent, segmented EBT translations. For example, the reference for PT is laera-quaresma and the reference for FR is noeismet. References are not translated by DeepL--they are simply shown in the output to aid in verification. The Pali MS segmented text is also shown in the output for an absolute reference of comparison.

To translate a sutta to portuguese, try this:

scripts/translate.mjs -dl pt an3.49

For a description of the options to translate.mjs:

scripts/translate.mjs --help

Glossaries and Transforms

Glossaries and transforms are located in src/glossary:

• Glossary files have the form ebt_SRCLANG_DSTLANG.kvg
• Transform files have the form transform_SRCLANG.json

Glossary

Glossary files are turned into DeepL Glossaries using the DeepLAPI. An uploaded glossaries will always replace the previous version. You'll always have a current glossary for your translations.

Transform files

Transform files are handled only by EBT-DeepL. Transform files are written in JSON:

{
  "Mönch oder eine Nonne": "Moench",
  "SRC_PATTERN": "REPLACEMENT"
}

Each entry in a transform file consists of a key/value pair. The key will be converted into a Javascript regular expression pattern. The value is the replacement.

If a transform file exists for a source language, all source text read by EBT-DeepL will be transformed by replacing text matching the regular expressions with their corresponding replacements.

A transform is useful when DeepL glossaries fail. For example, DeepL ignores "Mönch" glossary entries, but happily accepts and converts "Moench" glossary entries.

A transform is a useful hack that can steer DeepL when it feels stubborn.