Transforming dense, unstructured text into interpretable tables—commonly referred to as Text-to-Table generation—is a key task in information extraction. Existing methods often overlook what complex information to extract and how to infer it from text. We present Map&Make, a versatile approach that decomposes text into atomic propositions to infer latent schemas, which are then used to generate tables capturing both qualitative nuances and quantitative facts. We evaluate our method on three challenging datasets: Rotowire, known for its complex, multi-table schema; Livesum which requires numerical aggregation; and Wiki40 which require open text extraction from mulitple domains. By correcting hallucination errors in Rotowire, we also provide a cleaner benchmark. Our method shows significant gains in both accuracy and interpretability across comprehensive comparative and referenceless metrics. Finally, ablation studies highlight the key factors driving performance and validate the utility of our approach in structured summarization.

Text-to-table generation requires extracting structured information from unstructured narratives—a task that challenges current LLMs in determining what to extract and how to infer missing information. Map&Make addresses these challenges through a modular pipeline that scales test-time computation: atomizing text into propositions, extracting schemas, and generating tables.

This approach improves performance across datasets with different characteristics—from sports summaries requiring complex schemas to live commentary demanding numerical reasoning to open-domain content needing flexible extraction.

We employ a comprehensive evaluation suite combining String-Similarity metrics and Specialized metrics:

• String-Similarity Metrics: Exact Match, chrF (Popovic´, ACL 2015), and BERTScore (Zhang* et al., ICLR 2020) for cell-level accuracy against gold annotations. These metrics however are suitable to match table structures extracted attributes and player names (rows and column headers), they faily to capture discrepencies in cell level values.
• TabEval (Ramu et al., EMNLP 2024): Entailment-based evaluation that captures table semantics by comparing atomic statements, providing stronger correlation with human judgments.
• AutoQA:(Jain et al., ACL 2024) Question-answering fidelity to measure information coverage without requiring gold tables.
• Numerical Metrics: Error rates and RMSE for aggregation accuracy on Livesum.

Map&Make is a modular 3-step prompting pipeline that enables step-by-step reasoning:

This decomposition enables better information coverage and integration compared to end-to-end approaches, allowing models to reason explicitly about what to extract and how to structure it.

Beyond inference-time prompting, Map&Make serves as a scalable data generation engine. We used Gemini 2.0 with our pipeline to generate high-quality supervision on the RotoWire training set, then fine-tuned LLaMA 3 8B Instruct. Results show that fine-tuning with Map&Make-generated data yields clear gains over Chain-of-Thought prompting in AutoQA and player-level recall.

However, team-level recall drops (41.61 vs. 56.26), indicating challenges in aggregating broader content—likely due to formatting consistency and long-context reasoning limitations in smaller models. While the full Map&Make pipeline (inference-time) still outperforms both alternatives, these results demonstrate the framework's effectiveness as a data generation tool for low-resource and multilingual settings where fine-tuning is necessary.