Discourse-level Clause Complex Boundary Identification Based on Explainable Large Language Models

Accurate clause complex boundary identification is foundational for advanced discourse comprehension and deep semantic parsing. Existing end-to-end deep learning models achieve high performance but provide less interpretability. An interpretable recognition paradigm is proposed in this paper, combining syntactic features from the Language Technology Platform (LTP) with semantic features from a large language model (LLM). The approach employed a fusion decision algorithm to integrate LTP-extracted syntactic features and LLM-generated semantic labels in a weighted manner for joint boundary determination. Experimental results demonstrate the hybrid LTP-LLM model achieves an 88.24% accuracy, which is an effective balance between performance and interpretability.