Retrieve, Program, Repeat: Complex Knowledge Base Question Answering via Alternate Meta-learning

Knowledge base question answering (KBQA) has evolved from rule-based systems to neural models like Memory Networks and Transformer-based approaches. Early methods relied on template matching, but lacked flexibility. Recent advances include end-to-end neural models that directly map questions to logical forms, such as Seq2Seq and semantic parsing techniques. However, these models often struggle with complex, multi-hop, or diverse question types, limiting their generalization. Meta-learning approaches like MAML have been introduced to improve adaptability, but typically require extensive labeled data. Retrieval-based methods, such as S2A, incorporate similar question retrieval to personalize responses, yet face challenges in training efficiency and support set quality. This paper builds upon these foundations, integrating weak supervision and meta-RL to jointly optimize retrieval and program generation, aiming for scalable, accurate, and adaptable KBQA systems.

The core challenge in KBQA is enabling models to understand and generate accurate logical programs for a wide variety of complex questions, especially those unseen during training. Existing methods often depend heavily on annotated data, which is costly and limits scalability. Additionally, models trained as a single monolithic system lack flexibility, leading to poor performance on out-of-distribution questions. Retrieval-based approaches improve this by finding similar questions, but their effectiveness hinges on the quality of support sets. Moreover, traditional training paradigms do not effectively leverage the feedback from answer correctness to guide retrieval strategies, resulting in suboptimal support selection and slow adaptation. Addressing these issues requires a framework that can learn to retrieve relevant support dynamically, adapt quickly to new questions, and do so with minimal supervision.

This work introduces several key innovations: 1) joint training of retrieval and program generation models via meta-RL, enabling end-to-end adaptation; 2) weak supervision leveraging answer rewards to guide retrieval policy, reducing annotation costs; 3) a filter softmax mechanism that incorporates question type information, improving retrieval accuracy and efficiency; 4) a two-stage alternating training process that iteratively refines both components for optimal performance. These innovations collectively allow the system to rapidly adapt to unseen questions with minimal labeled data, outperforming prior methods that treat retrieval and generation separately or rely on extensive supervision. The integration of meta-learning with weak supervision marks a significant step forward in scalable, flexible KBQA.

• �� Construct a neural encoder-decoder (programmer) and a retriever network, both taking questions as input.
• �� Use a meta-RL approach: in the first stage, fix the retriever, sample support questions, and optimize the programmer via policy gradients based on answer rewards.
• �� In the second stage, fix the programmer, sample support sets using the retriever, and update the retriever policy to maximize the reward difference.
• �� Employ DSSM for semantic similarity, combined with a filter softmax to restrict candidate support questions based on question type.
• �� Use Monte Carlo policy gradient estimation to handle the non-differentiable support set sampling.
• �� Alternate these steps iteratively until convergence, ensuring both models improve jointly.
• �� During inference, adapt the program generator to new questions using retrieved support sets, guided by the learned policies.

The model was evaluated on the CQA dataset, containing 944K questions across seven categories. Baselines included KVmem and CIPITR, with metrics like micro and macro F1. The training used only 1% labeled data for pseudo-gold sequences, with additional unlabeled data for meta-learning. Hyperparameters included N=5 support questions, 30 training epochs, and Adam optimizer. Ablation studies confirmed the importance of joint training and the filtering mechanism. Results showed significant improvements over baselines, especially in complex reasoning categories, demonstrating the model’s ability to generalize and adapt with minimal supervision.

MARL achieved a micro F177.71%, outperforming KVmem (31.18%) and CIPITR variants (~60%). Ablation showed that removing the joint training reduced performance by nearly 3%. The support set quality directly correlated with accuracy gains, especially in multi-hop and reasoning-heavy questions. The model maintained robustness across categories, validating the effectiveness of the combined weak supervision and meta-RL approach.

This framework can be deployed in intelligent virtual assistants, enterprise knowledge systems, and customer support bots, where rapid adaptation to diverse queries is crucial. It requires a knowledge base, natural language understanding, and entity linking modules. Its ability to learn from minimal supervision makes it suitable for large-scale, real-world deployment, reducing annotation costs and improving user experience. Future integration with multi-modal data could further enhance its versatility.

Despite promising results, the model’s performance on highly complex, multi-hop, or ambiguous questions remains limited. The training process is computationally intensive due to alternating optimization and Monte Carlo sampling. Dependence on entity linking accuracy can lead to errors. Future work should focus on reducing computational costs, improving multi-hop reasoning, and enhancing robustness against noisy inputs.