神经符号模型将传统知识驱动的符号主义人工智能模型与最近高速发展的数据驱动深度学习模型相结合，试图改善当前深度学习模型不可解释和泛化性差的问题。神经符号模型通过生成有语义的符号结构解决各种推理任务，同时提供人类可理解的推理过程，增强其可信性。因此，符号结构生成过程对于神经符号模型来说至关重要。然而由于符号空间缺乏数据标注和直接的监督信号，已有神经符号模型的结构生成过程仍有较大改进空间。另外，神经符号模型在缺少先验知识引导时，不仅训练所需时间和计算资源更多，性能也较为落后，难以在复杂任务上实现高效推理。因此，结构生成和高效推理成为神经符号模型发展面临的两大关键挑战。本文研究工作的动机就是提出新方法应对这两项挑战，其主要内容包括以下四个方面：本文提出了一种基于对比学习的成对误差，借助数据集中的``互补对‘‘标注在原本缺乏直接标注的符号空间中引入自监督信号，从而提升神经符号模型结构生成的质量。本文首次将神经符号模型的结构生成过程建模为一个动态的马尔科夫决策过程，从而通过强化学习方法对结构生成器进行优化，解决已有模型中存在的梯度消失和依赖间接监督信号的问题。该方法也是首次将神经符号模型应用于解决抽象推理任务，在取得领先性能的同时充分发挥其可解释性和泛化性更强的优势。本文提出一种规则约束的贝叶斯符号化推理模型，将人类先验知识表示为可计算的约束规则，对贝叶斯结构生成过程进行后验约束，从而提高其推理性能和训练效率。本文针对具身问答任务中的零样本泛化问题，提出一种掩码场景图模型，通过场景图结构上迭代的消息传递过程，使模型自监督地学习对未见过物体的识别。与已有研究对比，该方法不依赖与特定任务强相关的先验知识，而是充分利用预训练跨模态模型中隐含的弱相关先验知识，实现了神经符号模型在复杂任务上的高效推理。

Neural-Symbolic Models combine traditional knowledge-driven symbolic AI models with the recent data-driven deep learning models. They aim to address the problems of lack of interpretability and poor generalization in current deep learning models. Neural-symbolic models can solve various reasoning tasks by generating semantically meaningful symbolic structures while providing human-interpretable reasoning processes, thus being more reliable. Therefore, the process of symbolic structure generation is crucial for neural-symbolic models. However, the structure generation process of existing neural-symbolic models is still challenging due to the lack of data annotations and direct supervision signals in the symbolic space. In addition, without guidance of prior knowledge, neural-symbolic models require more training time and computational resources to optimize, but they still hinder in performance. This makes it difficult to apply neural-symbolic models to achieve efficient reasoning on complex tasks, like embodied reasoning. Therefore, structure generation and efficient reasoning are the two key challenges to tackle for neural-symbolic models. The motivation of this thesis is to develop new methods to address the two challenges. Its main contributions include the following four aspects:We propose a pairwise training schema based on contrastive learning, which introduces a self-supervised signal into the symbolic structure space by utilizing "complementary pairs" in data and subsequently improves the quality of structure generation for neural-symbolic models. We first formulate the structure generation process of neural-symbolic models as a dynamic Markov decision process. By using reinforcement learning methods to optimize the structure generator, we address the problems of gradient vanishing and indirect supervision signals in existing models.We are the first to introduce the neural-symbolic model to tackle abstract reasoning problems and achieve leading performance, while fully leveraging the interpretability and generalization advantages.We propose a rule-constrained Bayesian symbolic reasoning model that represents human prior knowledge as computable constraint rules, in order to apply posterior regularization methods for the Bayesian structure generation process. With the introduction of prior knowledge, our model shows improved performance and training efficiency.We address the zero-shot generalization problem in embodied question answering by proposing a new masked scene graph modeling method, which learns to recognize unseen objects through iterative message passing on the scene graph structure in a self-supervision manner. Unlike existing research, our method does not rely on task-specific prior knowledge but fully utilizes weak implicit knowledge embedded in pre-trained cross-modal models, hence achieves efficient reasoning of neural-symbolic models on complex tasks.