Map&Make: Schema Guided Text to Table Generation

(ACL 2025) Map&Make is a schema-guided approach that decomposes text into atomic propositions to infer latent schemas, then generates tables that capture both qualitative nuances and quantitative facts. This repository contains prompts, evaluation scripts, and resources for reproducing and extending the work.

---

Overview

Map&Make addresses text-to-table generation: transforming dense, unstructured text into interpretable tables. The method:

1. Decomposes input text into atomic propositions
2. Infers latent schemas from these propositions
3. Generates tables (player and team statistics in RotoWire) that are faithful to the text

We evaluate on RotoWire (multi-table schema), Livesum (numerical aggregation), and Wiki40 (open-domain extraction). We also release a corrected RotoWire test set to address hallucination and missing-information issues in prior releases (see dataset README and paper for details).

---

Repository Structure

map-make/
├── code/
│   ├── evals/          # Evaluation scripts and metrics
│   ├── prompts/        # Prompts for Map&Make and baselines
│   └── README.md       # Code-level documentation
├── static/             # Website assets (images, CSS, JS)
├── index.html          # Project website
└── README.md           # This file

---

Prompts (code/prompts/)

Prompts are organized by dataset and by method (Map&Make vs. baselines).

Map&Make pipeline (RotoWire: prompts/Rotowire/)

File	Description
Atomization.txt	Decompose text into atomic, single-fact statements.
Schema_Extraction.txt	Infer table schema (headers, structure) from atomic statements.
Table_Generation.txt	Generate table rows from atomic statements under the inferred schema.
Unified.txt	Single prompt that chains Atomization → Schema Extraction → Table Generation (Unified CoT).
Unified_CoT.txt	Unified chain-of-thought variant.

These implement the pipeline described in the paper: text → atomic propositions → schema → tables.

Baselines (RotoWire: prompts/Baselines_Rotowire/)

File	Description
CoT_0shot.txt / CoT_1shot.txt	Chain-of-thought table generation (zero- and one-shot).
T3_Text_Tuple.txt	T3-style text-to-tuple formulation.
T3_Tuple_Table.txt	T3-style tuple-to-table.
T3_Tuple_integrate.txt	T3 integrated tuple+table.
T3_merged_1shot.txt	T3 merged one-shot prompt.

Used for comparison against Map&Make in the paper.

---

Evaluation (code/evals/)

Evaluation is split into cell-level metrics (EM, chrF, BERTScore), entailment-based metrics, and QA-based (AutoQA) metrics.

Cell-level metrics (em_chrf_bert/)

• em_chrf_bert_eval.py
  Evaluates generated tables against gold tables using:

  • Exact Match (EM) for headers and cells
  • chrF (character n-gram F-score)
  • BERTScore (e.g. RoBERTa-large) for semantic similarity

  Supports precision, recall, and F1 over headers and non-header cells, with optional grouping by row/column header.

• em_chrf_bert_collated.py
  Collates and aggregates results across runs or samples.

Table entailment (tabeval/)

• entailment.py
  Uses an NLI model (e.g. RoBERTa-large-MNLI) to check whether each gold proposition is entailed by the predicted table (unrolled to text). Computes precision, recall, and F1 over propositions (referenceless, table-as-output evaluation).

• tabunroll.txt
  Instructions/template for unrolling tables into natural language statements for entailment checking.

AutoQA (autoqa/)

• generate_qa.txt
  Instructions for generating fact-based questions from the source text (coverage of names, events, statistics, etc.).
• qa_on_tables.txt
  Instructions for answering those questions using the generated tables.
• check_answers.txt
  Instructions for checking QA answers against the text.

Used to measure whether the generated tables support the same factual questions as the source text (QA-based fidelity).

Utilities (utils/)

• eval.py
  Core table evaluation: parsing model output into tables, matching headers and cells to gold, and computing EM/chrF/BERTScore with configurable options.
• table_utils.py
  Helpers for loading, normalizing, and comparing table structures.
• utils.py
  General I/O and formatting utilities for evals.

---

Dataset

• Corrected RotoWire test set (text–table pairs)
  Hugging Face: McH04/Rotowire-Text-to-Table
  Clean benchmark for text-to-table; see the dataset card for motivation, structure, and citation.

• Livesum
  HKUST-KnowComp/LiveSum
  Used in the paper for settings requiring numerical aggregation.

---

Citation

If you use Map&Make or the corrected RotoWire dataset, please cite:

@inproceedings{ahuja-etal-2025-map,
    title = "Map{\&}Make: Schema Guided Text to Table Generation",
    author = "Ahuja, Naman and Bardoliya, Fenil and Baral, Chitta and Gupta, Vivek",
    booktitle = "Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
    month = jul,
    year = "2025",
    address = "Vienna, Austria",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2025.acl-long.1460/",
    doi = "10.18653/v1/2025.acl-long.1460",
    pages = "30249--30262"
}

---

License

This project is licensed under the MIT License — see the LICENSE file for details.

---

Links

• Paper: ACL 2025 | arXiv
• Project page: https://coral-lab-asu.github.io/map-make
• Dataset: Hugging Face – McH04/Rotowire-Text-to-Table