ODUTQA-MDC: A Task for Open-Domain Underspecified Tabular QA with Multi-turn Dialogue-based Clarification

Zhensheng Wang; ZhanTeng Lin; Wenmian Yang; Kun Zhou; Yiquan Zhang; Weijia Jia

cs.CLApr 11, 2026

ODUTQA-MDC: A Task for Open-Domain Underspecified Tabular QA with Multi-turn Dialogue-based Clarification

Authors: Zhensheng Wang, ZhanTeng Lin, Wenmian Yang, Kun Zhou, Yiquan Zhang, Weijia Jia

Organizations: School of Artificial Intelligence, Beijing Normal University, Beijing, PR China · Institute of Artificial Intelligence and Future Networks, Beijing Normal University, Zhuhai, PR China · Faculty of Arts and Sciences, Beijing Normal University, Zhuhai, PR China · Beijing Normal-Hong Kong Baptist University, Zhuhai, PR China

Abstract

The advancement of large language models (LLMs) has enhanced tabular question answering (Tabular QA), yet they struggle with open-domain queries exhibiting underspecified or uncertain expressions. To address this, we introduce the ODUTQA-MDC task and the first comprehensive benchmark to tackle it. This benchmark includes: (1) a large-scale ODUTQA dataset with 209 tables and 25,105 QA pairs; (2) a fine-grained labeling scheme for detailed evaluation; and (3) a dynamic clarification interface that simulates user feedback for interactive assessment. We also propose MAIC-TQA, a multi-agent framework that excels at detecting ambiguities, clarifying them through dialogue, and refining answers. Experiments validate our benchmark and framework, establishing them as a key resource for advancing conversational, underspecification-aware Tabular QA research.

Figures & tables

Figure 1: Examples of underspecified input causing questions to be unanswerable.

Dataset	Open Domain	Underspecification	Mixed Underspecification	QA Evaluation	Conversational	Dynamic Clarification
Spider	✗	✗	✗	✓	✗	✗
RETQA	✓	✗	✗	✓	✗	✗
PRACTIQ	✗	✓	✗	✗	✓	✗
AmbiQT	✓	✓	✗	✓	✗	✗
Ambrosia	✓	✓	✗	✗	✗	✗
ODUTQA	✓	✓	✓	✓	✓	✓

Table 1: Dataset comparison. “Open Domain” indicates table retrieval from a large corpus, whereas “Closed Domain” (marked with ‘✗’) specifies QA on given tables.

Figure 2: General framework of MAIC-TQA.

	Fixed Clarification							Dynamic Clarification
Model	SELECT		FROM		WHERE		Mixed	FROM		WHERE		Mixed
	Acc.	F1	Acc.	F1	Acc.	F1	Acc.	Acc.	F1	Acc.	F1	Acc.
BERT	99.78	99.22	–	–	–	–	–	–	–	–	–	–
Qwen3 32B	–	–	77.66	82.82	69.59	66.02	54.96	78.78	85.11	7 3.81	7 2.49	5 9.42
Qwen3 30B	–	–	75.17	85.10	75.67	78.99	5 8.55	69.95	77.45	72.49	69.50	52.92
Kimi K2	–	–	82.60	8 7.95	69.02	65.54	55.51	81.63	8 7.31	64.81	59.55	52.69

Table 2: Overall performance of underspecification detection under fixed and dynamic clarification settings.

Model	Method	Table Retrieval	SQL Generation
Model	Method	F1	ECR	EA
Qwen3 32B	SLUTQA	60.07	85.90	24.20
	Ours (Fix)	92.89	97.46	57.44
	Ours (Dyn)	93.48	98.13	57.89
Qwen3 30B	SLUTQA	58.78	86.75	21.22
	Ours (Fix)	95.10	96.32	54.11

Table 3: Table retrieval and SQL generation results (“Fix”/“Dyn”: Fixed/Dynamic clarification).

Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 3: Examples of Template Filling (Chinese–English Bilingual).

Figure 4: Illustration of an utterance with SLU tags. The example features two intents, where ‘B-C’ denotes ‘B-City’ and ‘B-Y’ denotes ‘B-Year’.

Figure 5: QA example.

Statistics	Num
Non-underspecified	5,095
SELECT underspecification	197
FROM underspecification	3,274
WHERE underspecification	4,421
Mixed-underspecification	12,118
Single intent	1,966

Appendix

Table 4: Dataset Statistics.

Model	Task	P	R	F1
BERT	Intent	98.51	98.61	98.56
BERT	Slots	99.68	99.88	99.78
Qwen3 30B	Intent	90.57	91.73	91.00
Qwen3 30B	Slots	97.01	96.79	96.90
Qwen3 32B	Intent	88.27	91.27	89.74
Qwen3 32B	Slots	96.07	97.23	96.64

Appendix

Table 5: Performance of Different Models on SLU Tasks.

Method	Model	ECR	EA
	Qwen3 32B	97.46	57.44
MAIC-TQA	Qwen3 30B	96.32	54.11
	Qwen3 32B	91.33	52.99
w/o SELECT	Qwen3 30B	92.63	49.30
	Qwen3 32B	80.54	35.98
w/o FROM	Qwen3 30B	81.01	31.84

Appendix

Table 6: Ablation study of the clarification modules in the MAIC-TQA framework.

Figure 6: Query rewriting prompts for dataset construction.

Figure 7: Prompt for sentence rewriting used in the dynamic clarification interface.

Figure 8: Prompt for slots classification.

Figure 9: Prompt for table caption summarization.

Figure 10: Prompt for SQL generation and condition underspecification detection.

Table QA

Table question answering studies how AI systems retrieve, interpret, and reason over structured data to answer natural-language questions.

See all 66 papers

TabScope: Question-Adaptive Scope Selection for Table Question Answering
Sep 3, 2026Yuxiang Wang, Junhao Gan, Jianzhong QiTable QA
The University of Melbourne
Conversational Task Disambiguation over Tabular Data: Leakage-Aware Formulation, Benchmark Suite, and Training
Oct 7, 2026Nafiseh Ghoroghchian, Luis Scoccola, Tina Sedaghat +4Task-Oriented Dialogue Systems Text-to-SQL
Layer 6 AI
Efficient Table QA via TableGrid Navigation and Progressive Inference Prompting
May 18, 2026Amritansh Maurya, Navjot Singh, Mohammed Javed +1Table QA Structured Prompting
Vision Intelligence Lab, IIIT Allahabad, Prayagraj, India · iMeDIA Lab, IIIT Allahabad, Prayagraj, India · CV:HCI Lab, Karlsruhe Institute of Technology, Karlsruhe, Germany

Sep 3, 2026J/K move · Enter open · S save

TabScope: Question-Adaptive Scope Selection for Table Question Answering

Yuxiang Wang, Junhao Gan, Jianzhong Qi

The University of Melbourne

Table QA

Open paper AskPDF arXiv

Large Language Models (LLMs) have shown strong performance on table question answering, yet their accuracy often degrades as table size increases. We find that this degradation is not uniform across question types. Localization-sensitive questions are particularly affected by irrelevant table content, while questions requiring broader evidence may still benefit from full-table reasoning. Based on this observation, we propose a question-adaptive framework that dynamically selects between localized and full-table reasoning. The framework constructs question-specific sub-tables through operation-aware table decomposition and uses the predicted question type to determine the appropriate reasoning mode. We further introduce silver reference sub-tables for evaluating evidence selection and construct SLQA, a benchmark based on real-world long tables. Experiments on WikiTQ and SLQA show that localization is particularly effective for lookup and local reasoning questions, while adaptive selection between localized and full-table reasoning achieves the best overall performance. These results highlight that long-table QA requires deciding not only how to localize, but also when to localize. Our code and datasets will be made available upon publication of the paper.

Abstract

Table QA

Explore similar work

TabScope: Question-Adaptive Scope Selection for Table Question Answering

Conversational Task Disambiguation over Tabular Data: Leakage-Aware Formulation, Benchmark Suite, and Training

Efficient Table QA via TableGrid Navigation and Progressive Inference Prompting