StructText
收藏资源简介:
StructText是一个端到端框架,用于自动生成从文本中提取键值的高保真度基准。它使用现有的表格数据作为结构化基准,并通过两阶段的“计划-执行”流程来合成生成相应的自然语言文本。为了确保文本与结构化源之间的对齐,我们引入了一种多维评估策略,该策略结合了(a)基于LLM的判断,涉及事实性、幻觉和连贯性,以及(b)测量数值和时间准确性的客观提取指标。我们在49个数据集的71,539个示例上评估了所提出的方法。结果表明,尽管LLM实现了强大的事实性准确性和避免了幻觉,但在生成可提取的文本方面,它们在叙事连贯性方面却遇到了困难。值得注意的是,模型以高保真度假设数值和时间信息,但这些信息嵌入在难以自动提取的叙事中。我们发布了一个框架,包括数据集、评估工具和基线提取系统,以支持持续的研究。我们的发现突出了一个关键差距:模型可以生成准确的文本,但在保持信息可访问性方面却遇到了困难,这是在各个领域实际部署的关键要求,既需要准确性,也需要机器可处理性。
StructText is an end-to-end framework for automatically generating high-fidelity benchmarks for key-value extraction from text. It leverages existing tabular data as structured benchmarks, and synthesizes corresponding natural language texts via a two-stage "plan-execute" workflow. To ensure alignment between generated texts and their structured sources, we introduce a multi-dimensional evaluation strategy that combines (a) LLM-based judgments covering factuality, hallucination, and coherence, and (b) objective extraction metrics measuring numerical and temporal accuracy. We evaluated the proposed method on 71,539 examples across 49 datasets. The results demonstrate that while LLMs achieve strong factual accuracy and avoid hallucinations, they struggle with narrative coherence when generating extractable texts. Notably, models generate numerical and temporal information with high fidelity, yet such information is embedded within narratives that are difficult to automatically extract. We release a framework encompassing the dataset, evaluation tools, and baseline extraction systems to support ongoing research. Our findings highlight a critical gap: while models can generate accurate texts, they face challenges in maintaining information accessibility—a key requirement for real-world deployment across various domains that demands both accuracy and machine processability.

- 1StructText: A Synthetic Table-to-Text Approach for Benchmark Generation with Multi-Dimensional EvaluationIBM Research · 2025年



