Go to ICLR 2026 Conference homepageDocs2Table: Bridging Multiple Documents and Structured Tables via LLMs
Keywords: Information Extraction, LLM, Table Generation, multiple document, Schema, Domain Identification, Sliding Window
TL;DR: We present a new multi-document to table generation task (Docs2Table) and the first multi-domain benchmark dataset FGLM, and designs a two-stage pipeline method DDST, achieving the state-of-the-art performance in cross-domain table generation.
Abstract: The task of Text-to-Table has garnered significant attention due to the critical value of structured data in information retrieval. Existing methods primarily focus on single-document scenarios, employing end-to-end generation to directly reproduce text fragments, while overlooking the practical need for comparative analysis of multiple documents and the issue of domain variability. To address this, we introduce a novel information extraction task—multiple documents to table generation(Docs2Table), which requires models to comprehend the content of multiple documents, identify their similarities and differences, and generate structured tables tailored to domain-specific needs. We construct the first multi-domain benchmark dataset in Docs2Table, FGLM, covering finance, government, law, and medicine, with all data sourced from real-world business scenarios. Furthermore, we propose a two-stage pipeline method named DDST (Docs-Domain-Schema-Table) that introduces a domain gate selection mechanism, integrates domain-specific characteristics, and leverages JSON Schema as an intermediary to generate table headers and their value constraints, thereby improving table generation accuracy. Experimental results demonstrate that DDST achieves state-of-the-art performance on both traditional datasets and FGLM, significantly outperforming existing methods, with further analysis indicating strong generalization capabilities across different domains.
Primary Area: datasets and benchmarks
Submission Number: 11342
Loading