Abstract:In text classification tasks, pre-trained language models have made remarkable progress in semantic representation. However, when modeling complex documents, they still suffer from limitations such as a single topic granularity and insufficient interaction between topics and context. To this end, this study proposes an adaptive hierarchical topic-aware ELECTRA neural topic model termed as HTA-ELECTRA. The model employs hierarchical variational topic modeling to jointly learn global and local topic representations, and introduces an adaptive threshold mechanism to dynamically select topics that are relevant to the current document. On this basis, a multi-level topic fusion strategy and a topic-aware attention mechanism are designed to progressively inject topic structure information into ELECTRA-based contextual representations, thereby enhancing document-level semantic modeling ability. Experimental results show that on the 20Newsgroups, MRD, and Yelp datasets, the proposed model outperforms multiple comparison models in terms of AUC, Macro-F1, Macro-Recall, Macro-Precision, and Accuracy, validating the effectiveness and robustness of the proposed method.