Words To Data - Convert Legal Documents Into Diffable Data Structures

Overview

words_to_data parses US Code titles and Public Laws (bills) from USLM XML format, providing structured access to legislative text, the ability to track changes between document versions, and tools for annotating how bills amend existing law.

Available for both Rust and Python with high-performance Rust core and ergonomic Python bindings via PyO3.

Features

• Parse USC and Public Law documents - Extract hierarchical structure from USLM XML files
• Rich text content - Capture heading, chapeau, proviso, content, and continuation fields
• Bill amendment extraction - Identify USC references and amending actions from bills
• Hierarchical diffing - Compute word-level differences between document versions
• Legal diff annotations - Link bill amendments to corresponding USC changes with verification status tracking
• Amendment similarity scoring - Calculate similarity between bill amendments for analysis
• Python bindings - Full API access from Python with PyO3

Installation

Rust

Add to your Cargo.toml:

[dependencies]
words-to-data = "0.2.0"

Python

pip install words-to-data

Note: Pre-built wheels are available for Linux x86_64. Other platforms will build from source (requires Rust toolchain).

Getting Data

• Title data: https://uscode.house.gov/download/download.shtml
• Bill data: https://congress.gov

Quick Start

Parse a US Code Document

Rust:

use words_to_data::uslm::parser::parse;

fn main() -> Result<(), Box< dyn std::error::Error>> {
    // Load a USCode Title
    let title_26 = parse("tests/test_data/usc/2025-07-18/usc26.xml", "2025-07-18")?;

    // Navigate to §174(a)
    let s174a = title_26.find("uscodedocument_26/title_26/subtitle_A/chapter_1/subchapter_B/part_VI/section_174/subsection_a").expect("§174 (a) not found");

    // Print the chapeau value
    println!(
        "§ 174(a) chapeau: {}",
        s174a.data.chapeau.clone().unwrap_or("<Empty>".to_string())
    );

    // Serialize
    words_to_data::utils::write_json_file(&title_26, "title_26.json")?;
    Ok(())
}

Python:

from words_to_data import parse_uslm_xml

title_26 = parse_uslm_xml("tests/test_data/usc/2025-07-18/usc26.xml", "2025-07-18")
s174a = title_26.find("uscodedocument_26/title_26/subtitle_A/chapter_1/subchapter_B/part_VI/section_174/subsection_a")
print(f"§ 174(a) chapeau: {s174a.data['chapeau']}")

Compute a Diff Between Versions

Rust:

use words_to_data::{diff::TreeDiff, uslm::parser::parse};

fn main() -> Result<(), Box< dyn std::error::Error>> {
    let doc_old = parse("tests/test_data/usc/2025-07-18/usc26.xml", "2025-07-18")?;
    let doc_new = parse("tests/test_data/usc/2025-07-30/usc26.xml", "2025-07-30")?;

    let diff = TreeDiff::from_elements(&doc_old, &doc_new);

    let s174a_diff = diff.find("uscodedocument_26/title_26/subtitle_A/chapter_1/subchapter_B/part_VI/section_174/subsection_a").expect("Section 174A has no changes, nor does its children!");

    for change in s174a_diff.changes.iter() {
        println!("{:#?} Changed:", change.field_name);
        println!("  Old: {}", change.old_value);
        println!("  New: {}", change.new_value);
        println!("  Number of word-level changes: {}", change.changes.len());
    }
    words_to_data::utils::write_json_file(&diff, "diff.json")?;
    Ok(())
}

Python:

from words_to_data import parse_uslm_xml, compute_diff

doc_old = parse_uslm_xml("tests/test_data/usc/2025-07-18/usc26.xml", "2025-07-18")
doc_new = parse_uslm_xml("tests/test_data/usc/2025-07-30/usc26.xml", "2025-07-30")

diff = compute_diff(doc_old, doc_new)

s174a_diff = diff.find("uscodedocument_26/title_26/subtitle_A/chapter_1/subchapter_B/part_VI/section_174/subsection_a")

for change in s174a_diff.changes:
    print(f"{change.field_name} Changed:")
    print(f"  Old: {change.old_value}")
    print(f"  New: {change.new_value}")
    print(f"  Number of word-level changes: {len(change.changes)}")

Extract Amendments from a Bill

Rust:

use words_to_data::uslm::bill_parser::parse_bill_amendments;

fn main() -> Result<(), Box< dyn std::error::Error>> {
    let data = parse_bill_amendments("tests/test_data/bills/hr-119-21.xml")?;

    println!(
        "Bill {}: {} amendments found",
        data.bill_id,
        data.amendments.len()
    );

    for amendment in &data.amendments {
        println!("Actions: {:?}", amendment.action_types);
        println!("Text: {}", &amendment.amending_text[..80.min(amendment.amending_text.len())]);
    }

    Ok(())
}

Python:

from words_to_data import parse_bill_amendments

data = parse_bill_amendments("tests/test_data/bills/hr-119-21.xml")

print(f"Bill {data.bill_id}: {len(data.amendments)} amendments found")

for amendment in data.amendments:
    print(f"Actions: {amendment.action_types}")
    print(f"Text: {amendment.amending_text[:80]}...")

Core Concepts

USLM Elements

Documents are represented as trees of USLMElement structures. Each element contains:

• ElementData: Metadata, text content, and identification
• Children: Nested child elements forming the document hierarchy

The library uses two types of paths:

1. Structural Path: Full hierarchy including all elements Example: uscodedocument_26/title_26/subtitle_A/chapter_1/section_174

2. USLM ID: Official USLM identifier (excludes structural-only elements) Example: /us/usc/t26/s174/a/1

Text Content Fields

Each element can contain up to five distinct text fields:

• Heading: Section or subsection title
• Chapeau: Opening text before enumerated items
• Proviso: Conditional or qualifying clauses
• Content: Main body text
• Continuation: Text appearing after child elements

Diffs

The TreeDiff structure mirrors the element hierarchy and tracks:

• Field changes: Word-level differences in text content fields
• Added elements: New child elements in the newer version
• Removed elements: Elements that existed in the older version
• Child diffs: Recursive diffs for matching child elements

Diffs are computed using word-level granularity via the similar crate.

Legal Diff Annotations

The LegalDiff structure wraps a TreeDiff and adds an annotation layer for linking code changes to their legislative source:

• ChangeAnnotation: Links one or more diff paths to a bill amendment
• BillReference: Identifies the bill and specific amendment text that caused a change
• AnnotationMetadata: Tracks verification status, confidence scores, annotator identity, and reasoning
• AnnotationStatus: Pending, Verified, Disputed, or Rejected

This enables building training datasets for ML models that predict how bills will modify existing law.

Amendment Actions

Bills can perform these operations on existing code:

Amend, Add, Delete, Insert, Redesignate, Repeal, Move, Strike, StrikeAndInsert

Annotator Tool

The annotator/ directory contains a Tauri desktop application for manually creating training datasets. It allows annotators to:

1. Load old and new USC XML versions alongside a bill
2. Select amendments from the bill
3. Highlight the specific text that causes each change
4. Link amendments to affected code sections
5. Export annotations as JSON

See annotator/README.md for setup instructions.

Note: The annotator is an early prototype and may change significantly.

API Documentation

Rust

Generate and view the full API documentation:

cargo doc --open

Development

# Run Rust tests
cargo test

# Build and install Python bindings locally
maturin develop

# Run Python tests
python -m pytest

License

MIT