Abstract: Answering natural language questions using information from tables (TableQA) is of considerable recent interest. In many applications, tables occur not in isolation, but embedded in, or linked to unstructured text. Often, a question is best answered by matching its parts to either table cell contents or unstructured text spans, and extracting answers from either source. This leads to a new space of TextTableQA problems that was introduced by the HybridQA dataset. Existing adaptations of table representation to transformer-based reading comprehension (RC) architectures fail to tackle the diverse modalities of the two representations through a single system. Training such systems is further challenged by the need for distant supervision. To reduce cognitive burden, training instances usually include just the question and answer, the latter matching multiple table rows and text passages. This leads to a noisy multi-instance training regime involving not only rows of the table, but also spans of linked text. We respond to these challenges by proposing MITQA, a new TextTableQA system that explicitly models the different but closely-related probability spaces of table row selection and text span selection. Our experiments indicate the superiority of our approach compared to recent baselines. The proposed method is currently at the top of the HybridQA leaderboard with a held out test set, achieving 21% absolute improvement on both EM and F1 scores over previous published results.
Introduction: The Rise of TextTableQA
Most modern Question Answering (QA) systems excel at processing either pure text or structured tables. However, real-world documents frequently present a hybrid context where tables are embedded within or linked to descriptive text passages. This paradigm, known as TextTableQA, requires models to reason across both modalities simultaneously. For example, a question might ask for a specific date mentioned only in a text passage linked to a specific row in a table.
The Challenge of Noisy Distant Supervision
The primary hurdle in training TextTableQA systems is the reliance on noisy distant supervision. In datasets like HybridQA, training instances often consist only of a (Question, Answer) pair. Because the "gold" answer string (e.g., "2018") may appear in multiple table rows or multiple text spans, standard training methods that treat all matches as correct introduce significant noise. This "bag" problem confuses models, as only one specific row or span typically provides the actual evidence required to answer the query.
The MITQA Architecture
To address these challenges, researchers from IBM and IIT Bombay introduced MITQA (Multi-Instance Table Question Answering). MITQA is a modular pipeline designed to denoise distant supervision and handle the memory constraints of Transformer models.
Key Modules:
- TableSplitter & PassageFilter: Tables are split into "retrieval units" consisting of a single row and its linked text. A query-informed PassageFilter then ranks and selects the most relevant text snippets to fit within token limits.
- RowRetriever with Multi-Instance Loss: Instead of a standard classification task, MITQA uses a multi-instance loss function. This allows the model to avoid a loss penalty as long as it assigns a high score to at least one correct row in a candidate bag, effectively denoising the retrieval process.
- AnswerExtractor with Multi-Span Training (MST): To handle multiple answer occurrences in text, MITQA employs a curriculum learning approach. It first trains on "easy" single-match cases and then uses that model to identify the most likely span in complex, multi-match cases.
- Joint Row+Span Reranker (RSR): This final stage combines confidence scores from both the retriever and extractor to select the most probable final answer.
Experimental Results
MITQA was evaluated on two major benchmarks: HybridQA and OTT-QA (Open Table-and-Text QA). On the HybridQA leaderboard, MITQA achieved a dominant position with an Exact Match (EM) of 65.5 and an F1 score of 72.7, marking a ~21% absolute improvement over previous baselines. In the open-domain OTT-QA setting, it achieved 36.4 EM, outperforming the previous state-of-the-art by approximately 4%.
Ablation studies confirmed that the Multi-Instance Loss and Joint Reranking strategies were the primary drivers of these performance gains, proving that sophisticated handling of supervision noise is critical for hybrid reasoning tasks.
Conclusion and Future Work
MITQA demonstrates that framing TextTableQA as a multi-instance learning problem significantly improves accuracy. While the current modular pipeline is highly effective, future research may explore end-to-end architectures to mitigate error propagation between modules. Additionally, extending the system's capabilities to handle complex numerical reasoning and specialized domains like finance or healthcare remains an important next step.
Summary updated