HybridQA: A Dataset of Multi-Hop Question Answering over Tabular and Textual Data

Question answering has evolved from simple fact retrieval to complex reasoning over multiple data sources. Early datasets like SQuAD focused on single passages, while HotpotQA introduced multi-hop reasoning within text. Concurrently, structured data sources like Freebase and WikiData enabled semantic parsing approaches such as SQLNet. However, these efforts often isolated structured and unstructured data, limiting real-world applicability. Recent advances aim to fuse heterogeneous information, but datasets supporting multi-hop, multi-modal reasoning remain scarce. HybridQA addresses this gap by providing a large-scale benchmark that combines tables and linked texts, fostering research into integrated reasoning models.

Existing QA systems struggle with multi-hop reasoning across heterogeneous data formats, often limited to single modality or shallow reasoning. The core challenge lies in effectively integrating structured tables with unstructured text, especially for complex questions requiring multiple inference steps. The lack of comprehensive datasets hampers progress, and current models fail to generalize well to real-world scenarios where information is dispersed across formats. Additionally, weak supervision signals introduce noise, complicating training. Overcoming these issues demands novel architectures capable of multi-source reasoning, robust training strategies, and scalable datasets.

This work's key innovations include: 1) constructing a large-scale hybrid dataset combining Wikipedia tables and linked texts, designed for multi-hop reasoning; 2) developing HYBRIDER, a multi-stage neural architecture that combines SQL parsing, passage retrieval, and reasoning modules; 3) implementing a multi-hop path matching mechanism to connect heterogeneous information sources; 4) employing weak supervision to train complex reasoning pathways efficiently. These innovations enable effective reasoning over complex, multi-modal questions, surpassing prior models limited to single data sources or shallow inference. The dataset and architecture set new standards for hybrid question answering research.

• �� Data collection: Crawled Wikipedia tables with 5-20 rows, 3-6 columns, extracted hyperlinked passages, and designed multi-hop questions requiring reasoning across table and text. • Model components: SQLNet-based table parser for structured reasoning, BERT-based retriever for textual passages, and HYBRIDER for multi-source fusion. • Training: Used weak supervision by aligning answers with source paths, trained modules independently to reduce complexity. • Inference: Sequentially performed ranking, hopping, and answer extraction, combining structured and unstructured information. • Evaluation: EM and F1 metrics compared across single-modality and hybrid models, validating the effectiveness of the fusion approach.

Experiments on 70K question-answer pairs demonstrated the superiority of the hybrid model, which achieved 52% EM and 58% F1, far exceeding table-only (12%) and passage-only (25%) baselines. Ablation studies confirmed the importance of multi-hop path matching and weak supervision training. The model maintained robustness across question types, including numerical, superlative, and multi-step inference. Hyperparameters such as retrieval threshold τ and BERT variants were tuned to optimize performance. The system was trained on GPU clusters, with training time around 24 hours, ensuring reproducibility and scalability.

Results clearly show that integrating heterogeneous information sources significantly boosts performance, especially on complex multi-hop questions. The fusion model's EM surpasses 50%, indicating effective reasoning capabilities. Ablation studies reveal that each component—retrieval, hopping, and answer extraction—contributes substantially to overall accuracy. The model demonstrates strong generalization across question types, including numerical comparisons and superlatives, validating the design choices. These findings highlight the potential of multi-modal, multi-hop architectures for real-world knowledge reasoning.

The system can be applied in intelligent virtual assistants, enterprise knowledge bases, and automated data analysis tools, where questions often span multiple data formats. It requires high-quality, linked multi-modal data and sufficient computational resources. The approach is suitable for domains like healthcare, finance, and legal analysis, where complex reasoning over heterogeneous data is essential. Future integration with knowledge graphs and reinforcement learning could further enhance its capabilities, enabling more autonomous and accurate reasoning in dynamic environments.

The model faces challenges in resolving conflicting information from heterogeneous sources, especially in ambiguous scenarios. High computational costs and reliance on large annotated datasets limit scalability. The weak supervision approach, while effective, introduces noise that may affect accuracy. Additionally, the system's interpretability and real-time performance need improvement for deployment in resource-constrained settings. Addressing these limitations will be critical for practical, widespread adoption.