PINGALA: Prosody-Aware Decoding for Sanskrit Poetry Generation

TL;DR

PINGALA introduces a prosody-aware decoding method for Sanskrit poetry generation, improving semantic coherence and metrical adherence by segmenting verses and using phonetic transliteration, with a new reference-free evaluation approach.

Contribution

The paper presents PINGALA, a novel decoding approach that enhances Sanskrit poetry generation by segmenting verses, incorporating phonetic transliteration, and employing cross-encoder based evaluation.

Findings

Semantic coherence improved by 10% with segmented lines.

Metrical alignment increased by 46% using phonetic transliteration.

Cross-encoder evaluation better aligns with true poetry instances.

Abstract

Poetry generation in Sanskrit typically requires the verse to be semantically coherent and adhere to strict prosodic rules. In Sanskrit prosody, every line of a verse is typically a fixed length sequence of syllables adhering to prescribed binary patterns of syllable weights. We observe that instead of treating a verse as a monolithic sequence, segmenting them as grouped-lines leads to significant improvement in semantic coherence by 10\% with comparable metrical adherence. Specifically, PINGALA, our proposed decoding approach is designed to encourage every line to have well-formed words and our token selection biases the model towards it by preferring longer tokens. Writing in Sanskrit follows phonemic orthography, hence using a phonetically aware transliteration scheme, SLP1, increased the metrical alignment by 46\% with comparable semantic similarity, for a instruction fine-tuned large language models like Phi-4. We also introduce a new approach for reference-free evaluation using cross-encoders which achieved better alignment with true poetry instances.