Accent Placement Models for Rigvedic Sanskrit Achieve Fidelity Using ByT5 and BiLSTM-CRF

The accurate reconstruction of ancient texts relies heavily on understanding subtle linguistic features, and the Rigveda, a foundational text of Indian culture, presents a unique challenge with its complex pitch-accent system. Akhil Rajeev P and Annarao Kulkarni, from the Centre for Development of Advanced Computing, lead a study that tackles this problem by developing computational models to automatically restore missing accent marks in Rigvedic Sanskrit. Their work establishes reproducible methods for accent placement, comparing the performance of a fully fine-tuned transformer model, a traditional sequence-labelling approach, and a parameter-efficient fine-tuning technique. By achieving high accuracy in accent restoration, this research not only advances the field of natural language processing but also unlocks new possibilities for digital scholarship, improved optical character recognition, and the synthesis of authentic Vedic chants.

Restoring Accents in Archaic Rigvedic Sanskrit

This study introduces a new framework for automatically restoring accents in Rigvedic Sanskrit, an ancient language presenting unique challenges for natural language processing. Scientists created a new dataset to evaluate accent restoration models, employing standard metrics like Word Error Rate and Character Error Rate, alongside a specialized Diacritic Error Rate to directly measure accent errors. They evaluated three models: ByT5, a powerful byte-to-byte transformer, LoRA, a parameter-efficient fine-tuning technique, and a more traditional BiLSTM-CRF sequence tagging model. The results demonstrate that modern NLP methods can be successfully applied to this challenging task, achieving reasonable accuracy in restoring accents.

ByT5 consistently outperformed the other models, suggesting its byte-level encoding and transformer architecture are well-suited to the task. LoRA provided a good balance between accuracy and computational efficiency, while the BiLSTM-CRF model proved less effective. This work establishes a foundation for future research, suggesting expansion of the dataset and exploration of alternative model architectures to further improve accent restoration accuracy.

Rigveda Accent Restoration Using Parallel Corpora

This research pioneers a new approach to restoring accents in the Rigveda by constructing a parallel corpus of over 22,000 aligned verse pairs. Each pair consists of an unaccented verse alongside its diacritically marked counterpart, providing a robust foundation for evaluating computational models. Scientists evaluated three models: full fine-tuning of the ByT5 transformer, a BiLSTM-CRF sequence labeler, and LoRA-based parameter-efficient tuning of ByT5. To assess performance, the researchers employed Word Error Rate, Character Error Rate, and a task-specific Diacritic Error Rate, isolating errors in graphemes and specifically measuring accent restoration accuracy. The results demonstrate that full ByT5 fine-tuning achieves the lowest error rates across all metrics, while LoRA offers a strong balance between accuracy and computational efficiency. The BiLSTM-CRF model provides a reproducible baseline for future research, establishing valuable resources for accent-aware Optical Character Recognition, Automatic Speech Recognition, and pedagogical applications in Vedic studies.

ByT5 Achieves High Accuracy Rigvedic Accent Restoration

This research presents a breakthrough in automatically restoring accents in Rigvedic Sanskrit, crucial for preserving oral tradition and philological research. Scientists constructed a parallel corpus of accented and unaccented verses and investigated three distinct strategies for automatic accent placement, utilizing ByT5, a byte-level Transformer model, and comparing full fine-tuning of this model against a baseline BiLSTM-CRF sequence labeler and a more efficient LoRA-based tuning method. The results show ByT5’s superior ability to accurately predict and restore the complex accent patterns of Rigvedic verse. LoRA-based tuning offered a strong balance between computational efficiency and accuracy, while the BiLSTM-CRF model served as a transparent and reproducible baseline. This research establishes reproducible baselines for Rigvedic accent restoration and provides valuable guidance for downstream tasks, including accent-aware Optical Character Recognition, Automatic Speech Recognition, and chant synthesis.

ByT5 Excels at Restoring Rigvedic Accents

This research presents a novel benchmark for automatically restoring accents in Rigvedic Sanskrit and demonstrates the successful application of modern natural language processing techniques to this challenging task. By evaluating three distinct approaches, full fine-tuning of the ByT5 model, a BiLSTM-CRF baseline, and LoRA-based parameter-efficient fine-tuning, the team established reproducible results and identified ByT5 as achieving the highest accuracy in accent restoration. The study highlights the importance of Unicode-safe preprocessing and mark-aware tokenization for effective processing of Vedic texts. The findings demonstrate that despite the limited size of available data and the unique constraints of the domain, contemporary NLP methods can be adapted to heritage language processing. This work lays the groundwork for systematic prosodic annotation of Vedic corpora, promising deeper linguistic analysis and a better understanding of ancient chanting practices.