Sebastian Nehrdich

Thrilled to share that I will be joining Tohoku University in Sendai, Japan, as tenure track assistant professor from autumn this year. Happy to work on AI for Sanskrit, Chinese, Tibetan, Pali, and Japanese, and to educate new generation of students who can work with these tools.

MITRA Explore is coming and it will be 🔥🔥

Coming very soon: MITRA Explore will enable to ask more open questions and get answers based on the powerful retrieval capabilities of Dharmamitra!

One recent observation by a colleague: “cracking complex Sanskrit sentences, understanding their grammar and translating them used to be a solitary joy. AI tools like dharmamitra make it less solitary, and less joyful.”

We are happy to announce that Dharmamitra now features a board of advisors. They will advise on the kind of data that Dharmamitra includes, on the functionality and design of our applications, and on making sure that we keep providing tools and utility of highest quality.

A bit late to the party but this is a great paper, and I am very happy to see that ByT5-Sanskrit is indeed a versatile model that adepts well and outperforms much larger LLMs in task-specific settings for Sanskrit!

This is sooo avoidable! Like there is zero intellectual effort in looking up the papers and citing them properly, at least that much effort one can expect right!

I am reviewing for various computer science conferences these days and I had to reject roughly 40% of the papers I looked at so far purely on the fact that their bibliographies are filled with hallucinations.

2. Mitrasamgraha: A Comprehensive Classical Sanskrit Machine Translation Dataset arxiv.org/abs/2601.07314 A large dataset of parallel sentence pairs for Classical/Vedic Sanskrit to English, covering multiple domains and time spans. Useful for machine translation!

Two preprints of the Dharmamitra project: 1. MITRA: arxiv.org/abs/2601.06400 This paper describes our large multilingual parallel dataset release, the machine translation model, and our retrieval system. 1/

Deep learning can of course be very helpful when it comes to the creation of such resources like annotated corpora, treebanks, etc. Thats where LLMs, POS taggers, etc. can come in extremely handy. I hope we will see more of this in 'mainstream Indology' and Buddhist Studies.

Linguistically transparent dependency treebanks, that enable the application of statistical modeling (think Bayes) are very different because traceable features can be defined, and developments can be modeled. 5/

An interesting observation is that people who work on 'the periphery' of Indology/Buddhist Studies (Vedic Sanskrit, or Tocharian, Kothanese etc.) are usually also much more likely to be trained in linguistics and to apply proper linguistic tooling. 1/

Exciting NLP doesn't necessarily require hardware that costs more than the yearly salary of a postdoc, if you have good data and the right ideas. Throwing more compute at a problem often helps, but sometimes going the opposite direction can lead to very elegant solutions.

It took some years for us to make transformer-based implementations truly competitive (i.e. aclanthology.org/2024.findings-…), but the 2018 model has kept the edge even during the LLM explosion of the recent years. 2/

It still amazes me that the data-driven Sanskrit word segmentation model Oliver Hellwig (with a bit of contribution of myself) in 2018 created (aclanthology.org/D18-1295/) was completely trained and evaluated on CPU within a couple of hours! 1/

At this year's SAT 令和大蔵経 event in Tokyo!

I gave my first public presentation here at Tohoku university yesterday! And we had the first snow in Sendai yesterday for this year. I am happy to be back in a country that has proper seasons…