Abstract
Hybrid Question-Answering (HQA) requires models to reason across both structured tables and unstructured text. We present TACR (Table-alignment-based Cell-selection and Reasoning), a novel framework that explicitly aligns natural language questions with table headers to improve evidence selection. By focusing on schema alignment before cell retrieval, TACR achieves state-of-the-art cell selection accuracy on the HybridQA benchmark and provides a more transparent, explainable reasoning process.
Introduction
Hybrid Question-Answering (HQA) is a "needle in a haystack" problem. In datasets like HybridQA, the answer to a question might be hidden in a table cell, or it might be tucked away in a Wikipedia passage linked to that cell. The challenge is two-fold: efficiency and explainability. It is computationally unrealistic to iterate over every row, column, and linked passage in a large table. Furthermore, many "black box" models fail to explain why they chose a specific piece of evidence, making them difficult to debug or trust in real-world applications.
Methodology
The authors propose TACR, which breaks the HQA task into a multi-stage pipeline focused on "alignment."
Table-Question Alignment
The core innovation of TACR is the explicit alignment of questions with table headers. Using an ALBERT-based encoder, the model learns to map specific words in a natural language question to corresponding table headers. For example, if a question asks about a "player," the model learns to focus on the "Athlete" column. This alignment layer allows the researchers to generate heatmaps showing exactly which words in a question triggered the selection of a specific column, providing a "paper trail" for the model's logic.
Cell Selection and Reasoning
Building on the Row-Column-Intersection (RCI) framework, TACR calculates the probability of each row and column containing the answer. By fusing these probabilities with the alignment data, it identifies the "top-k" candidate cells. Once identified, a RoBERTa-based QA module treats the row containing those cells as context, linearizing the row data and concatenating it with filtered text from linked passages to extract the final answer span.
Key Results
TACR was evaluated on HybridQA and WikiTableQuestions (WTQ). On the HybridQA dataset, TACR achieved a Hits@1 of 83.3% for cell retrieval, outperforming previous state-of-the-art models. The inclusion of the alignment module provided a significant performance boost, improving cell-selection accuracy by 2.5% to 4.3% across various metrics. On the full HybridQA task, TACR achieved an Exact Match (EM) score of 66.2%, outperforming most baselines and remaining competitive with massive models like GPT-3.
Conclusion
TACR demonstrates that aligning the question to the table schema is as important as finding the data itself. While the model currently struggles with complex numerical reasoning (e.g., aggregations), its alignment-based approach significantly narrows the search space and improves explainability. Future work will focus on developing stronger mathematical capabilities and scaling the approach to open-domain hybrid environments.
Summary updated