Knowledge-Augmented Deep Learning Framework for Chinese Literary Text Analysis Using ARO-HA-BiLSTM

Abstract

Chinese literary texts spanning classical, modern, and contemporary periods present significant challenges for computational interpretation due to their deep semantic structures, metaphorical richness, and historical context. Traditional NLP approaches often fail to capture hierarchical narrative dependencies and culturally embedded meanings. This study proposes a Knowledge-Augmented Literary Text Analysis Framework that integrates ChatGLM-based knowledge-enriched embeddings with a Hierarchical Attention-based Bidirectional Long Short-Term Memory (HA-Bi-LSTM) network, optimized using the Adaptive Remora Optimization (ARO) algorithm. A corpus of 2000 Chinese literary texts was pre-processed using tokenization, rule-guided normalization, and structure-aware segmentation to preserve linguistic and contextual integrity. The proposed ARO-HA-BiLSTM model achieved an accuracy of 0.9724, precision of 0.9851, recall of 0.9718, and F1-score of 0.9734, outperforming existing models such as EL-GCN (accuracy 0.8855) and TextCNN (accuracy 0.8613). Ablation analysis further confirms the effectiveness of knowledge-enriched embeddings and adaptive hyperparameter optimization in enhancing semantic and thematic classification. The results demonstrate that integrating knowledge- augmented Chinese large language models with hierarchical deep learning significantly improves cross- era literary interpretation. This framework provides a robust computational methodology for digital humanities research and supports scalable, machine-assisted analysis of culturally rich literary corpora.