ABSTRACT Objective The task of code generation aims to transform natural language descriptions into corresponding target code. Among the various approaches, syntax‐aware code generation has emerged as a significant approach that strives to generate code by directly modeling the underlying syntactic rules. However, existing works typically adopt an autoregressive approach to sequentially generate each abstract syntax rule, which inevitably neglects the rich structural semantic information inherent within the syntax rules. To address this issue, we propose an enhanced architecture of structure semantics based on Graph Neural Network for code generation. Methods Our approach explicitly models the internal structure of syntactic rules by treating them as graph data, thereby enabling the extraction of deeper structural semantics. Furthermore, we jointly model both the sequential semantics and structural semantics of syntactic rules, effectively addressing the limitations of solely sequence‐based approaches in capturing the inherent structural semantics of code. Results Experimental results on two widely used code generation datasets demonstrate that the proposed model consistently outperforms strong baselines, with gains of up to 2.14 BLEU points and 2.02 CodeBLEU points, highlighting the effectiveness of our structural‐semantic modeling approach for code generation.

Symbolic regression is a powerful technique for discovering mathematical expressions that best fit observed data.While neural symbolic regression methods based on large-scale pretrained models perform well on simple tasks, the reliance on fixed parametric knowledge typically limits their generalization to complex and diverse data distributions.To address this challenge, we propose a syntax-aware retrieval-augmented mechanism that leverages the syntactic structure of symbolic expressions to perform context-aware retrieval from a preconstructed token datastore during inference.This mechanism enables the model to incorporate highly relevant non-parametric prior information to assist in expression generation.Additionally, we design an entropy-based confidence network that dynamically adjusts the fusion strength between neural and retrieved components by estimating predictive uncertainty.Extensive experiments on multiple symbolic regression benchmarks demonstrate that the proposed method significantly outperforms representative baselines, validating the effectiveness of retrieval augmentation in enhancing the generalization performance of neural symbolic regression models.