A Bidirectional Hybrid Retrieval-AugmentedFramework for Yakshagana Text Simplificationand Semantic Retrieval

Yakshagana literature is composed in an archaic and highly poetic form of Kannada characterized by obsolete vocabulary, complex syntax, Sandhi–Samasa constructions, and culturally grounded expressions, making its interpretation difficult for contemporary Kannada readers. To bridge this linguistic gap, this work proposes a bidirectional framework comprising retrieval-augmented forward simplification and inverse semantic retrieval. In the forward Yakshagana-to-Simplification pipeline, semantically related Yakshagana–contemporary Kannada pairs are retrieved and supplied as in-context demonstrations to Qwen3-8B for generating meaning-preserving contemporary Kannada interpretations. In the inverse Simplification-to-Yakshagana pipeline, a contemporary Kannada description is used to directly retrieve the corresponding original Yakshagana verse without an LLM generation stage. Sparse BM25, dense cosine similarity, and hybrid sparse–dense retrieval are systematically evaluated in both directions using a domain-specific, human-verified parallel corpus developed for this study. Experimental results show that hybrid retrieval consistently provides the strongest performance. For forward simplification, the hybrid RAG configuration achieves 0.9512 Faithfulness, 0.8756 Answer Correctness, and 0.8912 Answer Relevancy. For inverse retrieval, the hybrid configuration achieves 0.8800 Context Precision, 0.9100 Context Recall, and 0.9400 Recall@3. The results demonstrate that combining lexical and semantic retrieval provides effective contextual grounding for Yakshagana-to-contemporary Kannada simplification while enabling accurate reverse retrieval of original Yakshagana verses from contemporary Kannada descriptions.