RnG-KBQA: Generation Augmented Iterative Ranking for Knowledge Base Question Answering

Knowledge Base Question Answering (KBQA) has evolved from rule-based systems like Lambda DCS to neural approaches leveraging pre-trained models such as BERT and T5. Early methods relied on handcrafted logical rules, which limited scalability and coverage. Recent advances include neural semantic parsers and ranking models that improve generalization but still face challenges in zero-shot and complex reasoning tasks. Despite progress, existing models struggle with unseen schema items, multi-hop reasoning, and noisy data, necessitating methods that combine semantic relevance with flexible generation. This research builds on these developments, aiming to enhance the ability of KBQA systems to handle unseen and complex queries effectively.

The main challenge lies in enabling KBQA models to accurately interpret and generate logical forms for questions involving unseen schema components or novel compositions. Traditional ranking models depend heavily on rule coverage, which is infeasible at scale, while pure generative models lack robustness in zero-shot scenarios. Addressing this requires a framework that can both semantically filter candidate logical forms and adaptively generate expressions for unseen patterns. Ensuring logical correctness, interpretability, and computational efficiency further complicates the problem, especially as knowledge graphs grow in size and complexity.

This paper introduces a hybrid framework combining a contrastive BERT-based logical form ranker with a T5 sequence-to-sequence generator. The ranker leverages semantic similarity learned via contrastive loss to select relevant candidates, overcoming coverage limitations. The generator then refines or constructs logical forms conditioned on the question and top-ranked candidates, enabling flexible adaptation to unseen schema. Negative sample bootstrapping enhances discriminative training, while execution-based validation guarantees logical correctness. This integrated approach outperforms prior methods, especially in zero-shot and complex reasoning scenarios, representing a significant leap forward in KBQA technology.

• �� Candidate logical forms are generated by graph search over the knowledge base, producing a diverse pool.
• �� The contrastive ranker, based on a BERT dual encoder, scores each candidate by concatenating question and logical form, trained to maximize the similarity of correct pairs and minimize that of negatives.
• �� Negative samples are mined via a bootstrapping strategy, where confusing false candidates are iteratively identified and used for training.
• �� The generator, built on T5, takes as input the question and top-k ranked candidates, producing a refined logical form through autoregressive decoding.
• �� During training, the model alternates between optimizing the ranker with contrastive loss and the generator with cross-entropy loss, using teacher forcing.
• �� In inference, beam search generates candidate logical forms, which are executed to verify validity; if none are executable, the top-ranked logical form from the ranker is used.

• �� Datasets include GrailQA, emphasizing zero-shot and compositional generalization, and WebQSP for standard QA.
• �� Metrics used are exact match (EM) and F1, with hyperparameters such as 3 epochs for the ranker, 10 for the generator, 96 negative samples, and Top-k=5.
• �� Baselines include pure ranking, pure generation, and previous state-of-the-art models.
• �� Ablation studies assess the impact of each component, negative sampling, and execution validation.
• �� The models are trained with Adam optimizer, learning rates of 1e-5 and 3e-5, respectively, with early stopping based on validation performance.

• �� On GrailQA, the full model achieves 68.8% EM and 74.4 F1, surpassing prior SOTA by over 10%, especially in zero-shot settings.
• �� On WebQSP, it attains 75.7 F1, outperforming previous bests like QGG (74.0), demonstrating robustness even without oracle entity linking.
• �� Ablation results show that removing either ranking or generation reduces performance significantly, confirming their synergy.
• �� Negative sampling improves discriminative ability, leading to better candidate filtering and logical form accuracy.

• �� The framework can be deployed in enterprise knowledge management, virtual assistants, and large-scale information retrieval systems, especially where unseen or complex queries are common.
• �� It supports multi-hop reasoning, compositional queries, and zero-shot generalization, making it suitable for dynamic knowledge environments.
• �� Future integration with multi-modal data and symbolic reasoning could further enhance interpretability and reasoning capabilities, broadening its industrial impact.

• �� The approach's effectiveness depends on the quality of candidate logical forms; extremely complex or noisy questions may still pose challenges.
• �� Computational costs are high due to multi-stage training, negative sampling, and execution validation, limiting real-time deployment.
• �� The logical expressions lack explicit interpretability and controllability, which are critical for explainability in sensitive applications.