MultiHiertt

A repackaging of the MultiHiertt dataset (Zhao et al., ACL 2022) for numerical reasoning over documents containing multiple hierarchical financial tables.

MultiHiertt is built on FinTabNet (CDLA-Permissive-1.0), extracting 4,791 multi-page documents from S&P 500 annual reports (1999-2019). Each document contains 2-6 hierarchical HTML tables and surrounding text. Questions require reasoning across multiple tables and/or text passages, with answers derived via arithmetic programs or direct span selection.

Why this version

The existing HuggingFace repository (yilunzhao/MultiHiertt) bundles model checkpoints alongside the dataset files, totalling 7.66 GB, and the dataset viewer is broken due to an Arrow conversion error on the HTML table fields. This version packages only the annotation data (172 MB) in a clean parquet format with a documented schema.

Splits

Split	Examples
train	7,830
validation	1,044

The reference-free leaderboard test set (1,566 examples, question only) is intentionally excluded.

Schema

Field	Type	Description
uid	string	Unique example ID (MD5 hash)
paragraphs	list[string]	Sentences from the document; table positions marked with ## Table N ## placeholders
tables	list[string]	Raw HTML string for each table in the document
table_description	string	JSON-serialized dict mapping "{table}-{row}-{col}" keys to natural-language cell descriptions (e.g. "Table 0 shows Revenue is $446.")
question	string	The financial question
answer	string	Numeric answer (e.g. "9805") or text span; coerced to string
program	string	Flat DSL reasoning program (e.g. add(10881,8729), divide(#0,const_2)); empty string for span-selection questions (~19% of train)
text_evidence	list[int]	Zero-indexed paragraph indices used as gold evidence
table_evidence	list[string]	Gold cell references in "{table}-{row}-{col}" format, matching table_description keys

table_description format

Keys follow the pattern "{table_idx}-{row}-{col}" (zero-indexed table, one-indexed row and column). Values are auto-generated natural-language descriptions of the cell in context, produced by the authors' preprocessing script:

import json
td = json.loads(row["table_description"])
# {"0-2-1": "Table 0 shows Revenue of 2007 is $446.", "0-2-2": ...}

Program format

Programs use the same DSL as FinQA: flat token sequences with #N back-references to prior step results and const_* predefined constants. Unlike FinQA, MultiHiertt does not include a steps decomposition or program_re nested form.

Usage

from datasets import load_dataset
import json

ds = load_dataset("rootsautomation/MultiHiertt")
ex = ds["train"][0]

print(ex["question"])
print(ex["answer"])
print(ex["program"])       # empty string if span-selection

# Reconstruct document with tables in position
for para in ex["paragraphs"]:
    if para.startswith("## Table"):
        idx = int(para.split()[2])
        print(f"[TABLE {idx}]: {ex['tables'][idx][:80]}...")
    else:
        print(para)

# Access cell-level evidence
td = json.loads(ex["table_description"])
for cell_ref in ex["table_evidence"]:
    print(f"  {cell_ref}: {td.get(cell_ref, '(no description)')}")

License

QA annotations and preprocessing code: MIT.

Underlying table data is sourced from FinTabNet, released under CDLA-Permissive-1.0. The original annual reports are publicly available SEC filings from S&P 500 companies.

Citation

@inproceedings{zhao-etal-2022-multihiertt,
    title     = "{M}ulti{H}iertt: Numerical Reasoning over Multi Hierarchical Tabular and Textual Data",
    author    = "Zhao, Yilun and Li, Yunxiang and Li, Chenying and Zhang, Rui",
    booktitle = "Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
    month     = may,
    year      = "2022",
    address   = "Dublin, Ireland",
    publisher = "Association for Computational Linguistics",
    url       = "https://aclanthology.org/2022.acl-long.454",
    pages     = "6588--6600",
}