Translating Buddhist Texts with AI

I am writing this post to refute some of the misconceptions that have arisen here on this forum and to give an overview of what we have been working on “behind the scenes”.

But first of all, I want to thank Bhante @Sujato for supporting women, LGBTIQA+, and BIPOC communities and supporting the research that is helping to undermine the patriarchal structure that has crept into our Sangha and other Buddhist institutions.

I initiated the translation of the Chinese Buddhist Texts with Artificial Intelligence together with the good people at DeepL. I needed these translations to write my article “Through the Yellow Gate” to gain an understanding of how discriminatory elements have entered the Buddhist texts and have created obstacles to the ordination and practice of LGBTIQA+ people. I am very happy I did this because it has positively affected the lives of real people; it has helped to inform monastic communities so more teachers are now open to the possibility of ordaining and accepting transgender and other LGBTIQA+ people.

It is impossible that all the Buddhist texts will be translated by humans in our very lives and I, therefore, see the use of AI as instrumental in helping us understand the history of the Buddhist words and to fight the discrimination that has crept into our Buddhist communities and the way the texts are explained for thousands of years (by men) under the guise of “It’s sacred, it’s what the Buddha said so we cannot touch it”. My research unearthed errors in human translation that have had a profound negative effect on groups of people up until this day.

I could not have done this research without two AI tools, both of which I have helped to develop: BuddhaNexus (together with @SebastianN at the University of Hamburg under the leadership of Dr. Orna Almogi) and the aforementioned Buddhist Chinese translations. Of course, such translations have to be used with caution; like Google Translate it’s far from perfect but it gives a general idea of what the text is about so the correct passages can be found and thereafter translated, and interpreted, by a human. The two tools together are especially powerful and help to find similar passages in different texts, which again aids the interpretation.

Sebastian, Marcus Bingenheimer, and Justin Brody came in later when Sebastian used the DeepL output to see if he could improve on the work. An announcement was published on H-Buddhism. This led to a large storm of criticism, a.o. by Meghan Howard. We have however since talked to many people with regards to their concerns and have listened to them. The result was that we didn’t publish our work on a website as we had initially planned but started working in a different direction; namely the creation of a translation engine like Google Translate or DeepL for Pali, Sanskrit, Buddhist Chinese and Tibetan. (Google Translate already has this for Sanskrit and this development has largely been hailed as very positive! It shows that people don’t take AI translations as gospel.) This project was called Linguae Dharmae.

Many things happened afterward as our project attracted the attention of Kurt Keutzer, AI specialist and professor of the Computer Sciences department at Berkeley University, and Sebastian was invited to work on AI translation models for ancient languages at the university. We had a meeting with Bhante @Sujato, Alex Wynne, and Kurt Keutzer about using the AI translations of the Pali commentarial texts for input into SuttaCentral’s Bilara system so Alex and his team could use this as a base for commentarial translations. (https://discourse.suttacentral.net/t/training-ai-models-on-the-suttas/27310/4) Unfortunately, this project fell through due to lack of funding.

During those initial phases working with this department in Berkeley, I decided to resign as I didn’t feel comfortable in this new setting as I felt it was too patriarchal*. But I also saw that the university’s expertise and resources could have an enormous impetus on the furthering of the AI translations of Buddhist texts, far more than we could have done at Hamburg University. Linguae Dharmae dissolved soon afterward.

At Berkeley University, Sebastian started another project without my involvement focussing on translating Tibetan Buddhist texts. This project was partly funded by the Dharamsala-based Monlam Foundation, an organization close to the Dalai Lama, and partly by Berkeley University. It is unfortunately only available in the Tibetan language and used intensively by Tibetan people around the world, serving up to 20,000 translations a day.

Recently, a new project has been started and Sebastian has asked me to join again. This project aims to provide more accurate machine translation from the four main Buddhist languages into English and several other languages (similar to Google Translate) as well as provide an alternative to BuddhaNexus in finding parallels between languages and corpora (i.e. unlike BuddhaNexus it will be able to find for instance parallels between Pali, Chinese, Sanskrit and Tibetan texts rather than just within one language). This project is hosted at the Berkeley AI Research Lab, University of Berkeley, but also consists of independent developers and volunteers and has established positive collaboration with other Buddhist institutions such as BDRC and Monlam.ai. I’m very happy that the team has many female and BIPOC people.

I hope this has created some more clarity and taken away some of the confusion and apprehension about the translation of Buddhist Texts with artificial intelligence. I see a lot of positive potential these tools can bring but like with everything, we need to be open to critical voices and adjust when needed; every tool can be used for negative purposes.

Even the Buddhist Texts themselves have been used for exploitation, discrimination, and especially the furthering of patriarchy in the Buddhist religion throughout history due to the work of human (male) translators and interpreters. Traditionally, Buddhist Studies has been a hobby of the aristocratic (mostly male) elites, which has fuelled the patriarchal status quo in the Buddhist Sangha. This AI technology is building a bridge so research can be done to learn the true meaning behind the Buddhist texts and not rely solely on the explanations of the privileged few.

I stand by what I have done and feel proud to have done it.

With metta
Ven. Vimala

* Just to clarify, this is a different department than the department that disinvited Bhante Sujato for the conference; which is something that did not involve me. I am not familiar with the Buddhist Studies department at Berkeley, only with the Computer Sciences department which has different managment.

16 Likes

Namo Buddhaya!

I think it’s good to make these tools as good as possible and it would aid in translation.

I am all for other people doing it if they want to, and i would use it to aid making my own master version of suttapitaka, when it’s available to me.

2 Likes

Thank you for posting this, Ayya.

I am currently using DeepL to translate Bhante @Sujato’s translations into other languages. My experience with DeepL is that it is, well, a bit random. Sometimes it translates A into X and sometimes it translates A into Y. So using DeepL in translation is a bit like playing that ancient game “Whack-a-mole”, where mischievous robot creatures would pop in and out of holes to avoid being hit with ab enthusiastically wielded padded hammer.

Using DeepL requires a lot of mindful attention. In fact, every single line of translated text must be critically reviewed meticulously and repeatedly. This is actually a lot of work, but it’s worth it since there are a lot of benefits to studying the Dhamma that closely. Another aspect of translating from one contemporary language to another is that Dhamma insights hidden by the source language can reveal themselves in the destination language.

DeepL cannot handle Dhamma translation on its own. People need to be involved extensively to review and edit the output. DeepL understands grammar and syntax. People understand the Dhamma, because people can understand suffering.

5 Likes

Thanks for sharing your perspective.

My concern with AI translation is that it may perpetuate and reinforce biased views and inaccurate translations. I am heartened that you are sensitive to this and hope much good will come out of your work.

Is it possible that they are different English words because they are different usages of the Pali and no English word conveys both Pali usages?

2 Likes

I think it’s been done… 42.

I believe AN8.54 can be of great use to see the true meaning behind the texts. It is about development, about guidance, about arriving somewhere (the other shore). Can’t we say that when this development does not happen in our lifes, the true meaning behind the texts is not yet really understood?

I feel, Dhamma is not difficult to understand but to make that step to really practice, to really give up all that heat, passions, fire, all that addiction to sense stimulation (especially also of the 6th sense domain) etc, that is so difficult.

I feel, in general, it is very easy to be here on this shore and pretent to be a great dhamma expert. Talk endlessly about the socalled true meaning of anatta, parinibbana, jhana, and impress many people with ones scriptural expertise, and meanwhile never set foot on the other shore or even step on the raft.

I feel it is very important to programm this in AI when we talk about true meaning behind the text.
Is that possible in AI?

Thank you for your observations and experiences @karl_lew .
I have changed the OP because I think I have been a little uncareful with my wording, giving the impression that the initial translation of the Chinese Buddhist Texts corpus was done with DeepL as you also see it on their website. It was not. The AI model was a cooperation between me and the people behind DeepL (as they have a lot of experience). The model was trained on very scarce translations of Buddhist Chinese. Buddhist Chinese is not the same as modern Chinese. In fact, there is not even one “Buddhist Chinese” because Chinese has developed as a language over thousands of years. So it’s very tricky, and the output of this model was just a first step, not suitable to be used by DeepL on their website. I am sure the people behind DeepL would not want us to conflate that model with the models for modern translation that are used on their website. So my apologies for the confusion.

Saying that however, you bring up a good point. No AI translation will ever replace human translation. They are a tool that can be used (like you do now) by people wanting to translate something. These tools exist and we use them, but we know that the output is never perfect. I still remember when these tools started and the outcome of it was often laughable! Nobody took it seriously. Now the models have improved and tools like these are used for modern translation all over the world. Certainly not perfect yet but improving.

With the current Buddhist texts, we have a lot of human translations where errors have crept in that have reinforced biased views. This is one thing I found out when writing the article I mentioned in the OP: that a lot of misconceptions are based on the wrong translations and interpretations that crept in nearly 2000 years ago that indeed have reinforced biased views. The original texts in the original language went to China, and were translated and the wrong translations were later been picked up by Rhys Davids when he started making a dictionary for Pali and those wrong translations persist until today, having negative effects on many people.

So yes, you are right, this danger always exists. I think we have to remember that AI is made by humans so it can never be better than what humans put into it. So we have to train it, like a little child, to use more accurate translations. That’s why it Bhante Sujato’s translations as training material are so very important! This can make the difference between those wrong translations persisting or being rectified.

Absolutely! Pali and English words are not always translatable one-on-one. A word can have a meaning that is not easily conveyed in another language and for which no one word exists. That is why translators often discuss and argue about the meaning of certain words. That discussion is very important so as to understand the true meaning of the word. But every translator has to choose one word that for them conveys the meaning best according to their understanding. That is why you read f.i. a translation by Bhante Sujato and he uses different words from Bhikkhu Bodhi. Sometimes it is good to study different translations of the suttas to be able to understand the true meaning.

I wish … I estimate we still have 80% to go …
At least we have good translations now for the Early Buddhist Suttas and Vinaya in English …

That’s what it is all about. We shouldn’t forget that we have to do what the Suttas point to, not merely always study them (or translate them). The true meaning becomes clear through practice, but study (including translations) and discussion are important to give us guidance and inspiration.

It is important to not throw all AI projects on one heap. There are many types of AI and we use it every day, even though we might not always see it. My Grammarly AI for instance tells me you made a spelling mistake in the above paragraph. Google Translate can translate your paragraph into another language. ChapGPT can write a poem for you …

BuddhaNexus finds matches between the various texts in Pali, Sanskrit, Tibetan, and Chinese. And what we are working on at the moment is something that does the same but also finds matches with passages in the other languages rather than just within the same language and provides an English translation. However the user must be clear when something is a human translation and when something is an AI translation. It is very important to put that in!

6 Likes

Namo Buddhaya!

I can add a couple points.

As i understand it, there is probably way too little pali translations around to train an AI, if one takes all dictionaries and translations it’s going to be a mess for two reasons

  1. Much variance in translations to begin with.
  2. Not much volume compared to modern languages.

I want it to be able to do two things more or less reliably

  1. Paraphrase the existant translations and change certain terms.
  2. Make new translations of untranslated text

I think #2 is probably impossible but maybe #1 can be made.

The current models have become better at understanding and predicting Pali also with the amount of data we have, even though it is of course far less than for modern languages.

Indeed Pali has evolved over centuries and the commentaries use language different than the Early Suttas. That is why I find it very sad that the Pali translation project with Bhante Sujato and Alex Wynne of Oxford University did not get off the ground due to a lack of funding. It would have been great to be able to get those translations. We would be able to feed the Bilara system with the AI translations and then the human translators could look at that and use that as an initial suggestion and with every human input, the AI system would learn to improve again.

4 Likes

I don’t mean to quibble, but the free version of ChatGPT isn’t giving answers based on retrieving information from the internet. It’s giving plausible (but possibly completely inaccurate) accurate sounding sentences. It has no understanding of the words it is spitting out, just that statistically they go together.

This to me is the real problem of using these LLMs for translation in the wrong hands (read: the public). They are able to sound like natural English and they will prioritize sounding natural over not. At least with the older language translation tools it was obvious that the tool was not able to handle the job it was given.

All that said, I believe that you, Ayya, have used the tools in a perfectly reasonable and acceptable way, not that’s it’s up to me to decide. The kind of research you are doing is very difficult considering the lack of translations. Unfortunately, though, I don’t think the research will be complete until there is a complete human translated canon available. But that doesn’t reduce the value of your work or mean that you shouldn’t be undertaking it.

4 Likes

@Snowbird - You are right, and I have changed the post. I have looked at ChapGPT exactly twice in my life, both times in the last few months just to see what it would do. So I never studied it very well but it is also not something I find very helpful for anything much.

It is not possible that humans will translate all the Buddhist texts in this lifetime, but AI can find relevant passages and give enough of a translation so we understand what the passage is most likely about. Then you can focus on translating just those suttas or passages that are meaningful for one’s current research. As I mentioned above, it is also important that the user is made aware that this is a machine-translated passage.

I disagree. The research will never be complete. Every human translation is an interpretation of the text and interpretations between humans differ and unfortunately often also reflect current societal influences like patriarchy. In the history of the Buddhist texts, it becomes very clear that the further you go in history the more there are changes in the stance towards women and LGBTIQA+. I would never say that just because a human has translated it, it is accurate! Far from.

7 Likes

What a fascinating project! Is it really dead? :grimacing:

One of the most useful features of DeepL is its presentation of alternate phrasing of translations. This critical feature is quite important to translators since it provides a locus of meaning rather than the anemic single result offered by Google Translate. Having that capability with Pali would be sooooo helpful. I’ve been using three Pali dictionaries to get the same effect and it is quite painful (EN/ES/PT).

I would be very interested in engaging with such a project.

I always feel horrible paraphrasing the Dhamma because it flattens a multi-dimensional understanding into spoon-fed pablum. Yes, the Dhamma is holographic, but that doesn’t mean we should smash it up and blender-ize it. The issue I have with mindless use of AI is that the result is Automated Idiocy, the other meaning of AI.

2 Likes

My hypothesis for the systematic use of synonyms throughout the canon is twofold.

The first consideration is “error correction”. To explain this, just look at any credit card. Credit cards have more digits than are required to identify a person. With 8,000,000,000 people, you don’t need 16+3 digits. You just need 10. Error correction preserves meaning in the face of random corruption. The canon uses repetition and synonyms in a marvelously rigorous way that effectively provides error correction. I experienced this “error correction” as studying the canon eventually changed my own native definitions of words to match the consistency of the Dhamma. Studying the Dhamma changed my vocabulary and thereby my view of the world.

The second consideration is “scope of meaning”. Synonyms reveal the facets of a concept, those subtle variations in usage that allow us to communicate nuances with precision and grace. Synonyms let us share a rich manifold of experience with one another.

As you say, the Pali and English words are not one-to-one, but the clusters are one-to-one as near as language permits. It does get difficult when one language has 5 words and another language has 4 words, however. :laughing:

So as an engineer looking at the canon, I’m absolutely gobsmacked with awe at the architecture of the canon and with its ability to sail the seas of time to deliver its wonders here and now.

4 Likes

Namo Buddhaya!

I remember being taken back thinking ‘okay so you are saying ‘wisdom’ is a faculty & a power, to be developed, and it is thought to be conjoined with conciousness in a way that a difference can be delineated but not a separation, this is quite extraordinary, okay tell me more’

1 Like

Perhaps in intermediate AI could have annotated my “42” with the appropriate context… maybe

Or such depending on bias.

That would be an awesome use of AI.

In a nutshell, the premis of Adams’s classic work is;
the earth itself is an super computer - designed by the universes most evolved beings, white mice (note the irony) - to answer the ultimate question. Which answer turns out to be… 42. Mice have to then go back to the design phase to build an even better computer to formulate an appropriate question.

My point is this:

to learn the true meaning behind the Buddhist texts

is quite a different goal from universal translation.

“I teach the end of suffering.” He stated that he had succeeded in doing that completely and I’m confident of those words. It’s now our job to apply the clearly communicated teaching.

Assuming (I’m not saying you do, but it sounds that way) there is an esoteric truth hidden by a conspiracy of the corrupt seems to me a distraction from the work plainly laid out for us… and, if history is too be respected, a precursor of cultic thinking and behavior. I trust I read you incorrectly, and hope you tolerate my mistakes with compassion.

The grain of irritating sand (suffering) that creates the pearl (release) is unique to each of us. Yours may be oppressive patriarchy. Mine may be unrelenting physical pain.

The Buddha teaches how to escape this veil of tears, not how to fix it. I hope your translation work goes well and that release becomes accessible to all beings.

Perhaps I’ve misunderstood, but I am inclined to think that the Buddha relates these two here:

AN1.114:1.2: Negligence leads to the decline and disappearance of the true teaching.”

A translator, human or AI, who has not learned the true meaning behind the Buddhist texts is negligent to the extent that they have not learned the true meaning behind the Buddhist texts.

Clear communication is actually quite difficult.

1 Like