Computational Approaches to Buddhist Philology: The Intellexus Project

Intellexus is an ERC Synergy Grant project that brings together philologists at the University of Hamburg and NLP researchers at Reichman University to build computational tools for the vast Sanskrit and Tibetan Buddhist corpora. This panel presents four computational works from the project. Classical Sanskrit and Tibetan are low-resource languages with rich morphology, non-Latin scripts, and little annotated data, so standard NLP tools work poorly on them. The four presentations introduce benchmarks, models, and annotation pipelines built specifically for these languages.

- DharmaBench: Evaluating Language Models on Buddhist Texts in Sanskrit and Tibetan. The first multitask benchmark for these traditions, with 13 classification and detection tasks drawn from canonical and commentarial sources, used to test a range of language models from proprietary systems to domain-specific open-weight models. - Do We Really Need So Many Tokens? The Performance–Efficiency Trade-off in Tibetan and Sanskrit. Vocabulary expansion for adapting LLMs to Tibetan and Sanskrit shows a clear efficiency–performance trade-off: small expansions (~64 tokens) give the best downstream performance, while larger expansions improve efficiency but hurt performance under limited continued pre-training data. - Automatic Segmentation of Classical Tibetan Texts into Autochthonous and Allochthonous Regions. Fine-tuned multilingual encoders separate indigenous Tibetan compositions (autochthonous) from translated Indic material (allochthonous), framed as a boundary-detection task and supported by the first annotated corpus released for it. - Scaling Sentence Similarity for Classical Tibetan with Automatic Annotations. A pipeline combining contrastive bootstrapping, iterative pair mining, and automatic annotation trains sentence embedding models for Classical Tibetan, reaching state-of-the-art similarity performance without large expert-annotated datasets and enabling semantic search across the corpus.