Work

T2CycleIE: Treating Hybrid Multi-Hop Table-Text Question Answering as Cycle Information Extraction

Year: 2026

Type: article

Abstract: Table-text question answering (TableTextQA) task, which concerns multi-hop reasoning over hybrid tabular and textual evidence based on the given questions, has witnessed significant advancements. Existing methods usually process tables and text as two isolated modalities or implicitly leverage the weak semantic connections between them. However, these methods ignore the essential fact that tabular and textual evidence are explicitly and strongly connected through entities they target, which results in severe retrieval omission and reasoning bias. To address these issues, in this paper, we transform hybrid multi-hop TableTextQA into a cycle information extraction task centered on entities and propose a fully zero-shot unified framework called T2CycleIE, which consists of three modules orchestrated by an entity-oriented cycle strategy to effectively retrieve and reason over connected hybrid evidence. Specifically, T2CycleIE reformulates TableTextQA as a cycle of entity-aware retrieval (EAR), local triple reasoning (LTR), and global alignment reflection (GAR), where these three modules are responsible for retrieving tabular and textual evidence via shared entities, reasoning over extracted local entity-centric target triples, and aligning hybrid target triples for reflective reasoning, respectively. Furthermore, inspired by the step-by-step reasoning behavior of humans, we design an entity-oriented cycle strategy based on large language models (LLMs) to collaboratively and cyclically orchestrate the EAR, LTR, and GAR modules for progressive retrieval and reasoning, thereby mitigating retrieval omission and reasoning bias issues. Extensive experiments on the HybridQA and OTTQA benchmark datasets demonstrate that our zero-shot method performs better than other heavily engineered state-of-the-art baselines.

Source: IEEE Transactions on Audio Speech and Language Processing

DOI: 10.1109/taslpro.2026.3739173

Authors: Shu-Xun Yang, Zhijing Wu, Heyan Huang, Xian-Ling Mao

Institution: Beijing Institute of Technology

Language: English

Topic: Topic Modeling

Subfield: Artificial Intelligence

Field: Computer Science

Domain: Physical Sciences

Open Access status: hybrid

Estimated APC paid: $2,645

Funders: National Natural Science Foundation of China, Beijing Institute of Technology Research Fund Program for Young Scholars, National Key Research and Development Program of China