<b>Bee-RAG for Auditable Japanese-Mandarin Chinese Sentence Alignment with Retrieval Evidence and Multi-objective Calibration</b>

Existing approaches often rely primarily on direct semantic similarity or batch-level neighborhood correction, making it difficult to jointly exploit lightweight task adaptation, bilingual retrieval evidence, selective review, and provenance-preserving output within a unified CPU-efficient pipeline.

This study proposes an auditable sentence-alignment framework that integrates a frozen multilingual MiniLM encoder, projection-level contrastive adaptation, retrieval-derived bilingual evidence, and multiobjective bee-colony calibration.</p>