Neuro-Symbolic Sentence Ranking for Malayalam Extractive Summarization with Conservative Dynamic MMR

Malayalam remains comparatively under-resourced for sentence-level summarization, while its agglutinative morphology, complex orthography, and news-writing conventions create additional modeling and deployment challenges. This paper presents a resource-conscious extractive summarization pipeline that ranks complete Malayalam source sentences instead of generating new text. The system evolved from an unsupervised LaBSE centroid baseline to a supervised sentence classifier and finally to a dual-path neuro-symbolic architecture. The semantic branch consumes a frozen multilingual sentence embedding; the symbolic branch encodes sentence position, normalized length, complex-word density, and numeral density. Their fused representation estimates sentence salience. A refactored Dynamic Maximal Marginal Relevance (D-MMR) selector normalizes relevance and cosine-similarity scales, bounds heuristic contributions, and protects classifier ranking through relevance-retention checks and fallback behavior. Five checkpoint and encoder variants were compared on a 50-article four-column consensus development set under oracle-length and production-length conditions. The Chotta Bheem checkpoint achieved the highest observed oracle-length Micro F1 of 0.6241 and Macro F1 of 0.6200, and the highest production-length Micro F1 of 0.5141. Chotta Bheem V2 remained qualitatively useful on broader frontend examples but did not improve aggregate sentence-index overlap. The results establish Chotta Bheem as the strongest current deployment checkpoint while also showing that the development set is not a substitute for a new, document-disjoint human benchmark.