A lightweight text chunking method fusing weakly supervised distillation and discourse relation discrimination

The performance of Retrieval-Augmented Generation (RAG) is highly dependent on the quality of text chunking in its knowledge base. For logically structured long documents such as textbooks and social science literature, existing chunking methods struggle to balance semantic integrity with boundary accuracy, often fragmenting coherent arguments and undermining the quality of retrieved context. This paper proposes a lightweight text chunking method, DH (Distilled- Hybrid), which fuses weakly supervised distillation with discourse relation discrimination to address this challenge. The method transfers the powerful implicit semantic boundary detection capabilities of a Large Language Model (LLM) to a lightweight student model via knowledge distillation. Concurrently, it innovatively designs a Discourse Relation Classifier (DRC) to explicitly capture logical connections between sentences. The core hypothesis is that these implicit semantic and explicit logical signals are highly complementary. A small fusion network jointly leverages them to make a unified decision, ensuring semantic integrity while precisely maintaining the logical continuity of the text. Experiments on a dataset from the textbook Outline of Modern Chinese History show that our method improves the boundary prediction consistency metric (𝐹1@ ± 1) by 16.5% over a semantic similarity baseline and significantly enhances downstream retrieval performance, achieving notable gains on key metrics such as Recall@5 and nDCG@5. The method achieves an inference speed of under 50ms per thousand characters on a CPU, presenting a solution that effectively balances chunking effectiveness and computational efficiency.