Running summarizing tools on a new article
Check the paper below, then start its synthesis.
Lightweight AST-Fused TextCNN
Zhiqiang Lyu (Beijing University of Technology, Beijing Polytechnic University)
Abstract: To meet the low-cost educational AI requirement of ICAITA 2026 Track 1, this paper proposes CodeAware-TextCNN, a zero-GPU lightweight model for closed-set programming error classification on hybrid text-code student queries, clearly distinguished from intent classification, knowledge retrieval and generative LLM tutoring. Existing LLMs and transformer-based code pre-training models demand GPU resources and incur heavy $\mathbf{O}\left(\mathbf{N}^{\mathbf{2}}\right)$ self-attention overhead, while vanilla CNNs ignore hierarchical AST syntax and misclassify token-identical queries with distinct syntax bugs. Our model extracts lexical GloVe embeddings, dual-branch AST structural features and TF-IDF global semantics, fused via low-complexity linear sigmoid gating to eliminate transformer computational bottlenecks. We build a 1,960 manually annotated undergraduate Q&A dataset with interannotator Cohen’s Kappa $\kappa=0.87$. CPU-only experiments show our method achieves 89.2% Accuracy and 87.5% Macro-F1, matching RoBERTa-base; it delivers 6.2 ms average latency (9.8 ms P95) and reduces model volume by 98% compared with. Ablation, dataset saturation and multi-user concurrency tests validate AST structural priors significantly boost shallow CNN performance for educational code tasks. Supplementary crosscourse and cross-language transfer experiments verify strong generalization across unseen courses and alternative programming languages. This lightweight architecture meets sub-10 ms real-time inference demands of campus PCs, providing a low-cost tutoring AI for resource-limited universities.