A Multilingual Human Annotated Corpus of Original and Easy-to-Read Texts to Support Access to Democratic Participatory Processes

This paper introduces a freely accessible, multilingual corpus of original and human-annotated Easy-to-Read texts in Spanish, Catalan, and Italian, designed to support automatic text simplification research and enhance access to democratic participatory processes.

The Gist

Imagine you are trying to explain a complex recipe for a gourmet dish to a friend who has never cooked before. If you use words like "sauté," "deglaze," and "reduction," they might get lost. But if you say, "fry the onions until golden," "add the wine and let it bubble away," and "cook until thick," they can follow along easily.

This paper is about creating a massive, multilingual "cookbook" for the internet, but instead of recipes, it's about democracy.

Here is the story of the paper, broken down simply:

1. The Problem: The "Wall of Text"

Imagine democracy as a giant town hall meeting. Everyone is invited to vote, debate, and make decisions. But the papers handed out at the meeting are written in "legal-speak" or "academic-speak"—dense, confusing, and full of big words.

For many people—those with learning disabilities, low literacy, or who are new to the country—this "Wall of Text" is a locked door. They can't get in. They can't participate because they can't understand the rules. The authors of this paper say: "We need to knock down that wall and build a ramp instead."

2. The Solution: "Easy-to-Read" (E2R)

The team created a special type of writing called Easy-to-Read (E2R). Think of this as translating a dense, boring legal document into a friendly, clear conversation.

• Short sentences: Instead of a paragraph that goes on for five lines, they break it into bite-sized chunks.
• Simple words: Instead of "rectify," they use "fix." Instead of "discrimination," they use "being treated badly."
• Clear visuals: Using bullet points and clear headings, like a menu at a restaurant.

3. The Big Challenge: The "Three-Legged Stool"

Usually, when people want to teach computers to do this simplification, they need a huge library of examples: the original hard text and the easy version side-by-side.

• The Gap: For English, there are some books in this library. But for Spanish, Catalan, and Italian, the library was almost empty. It was like trying to teach a robot to speak three different languages when you only had a dictionary for one.
• The Fix: The authors (a team from universities and disability organizations) decided to build this library from scratch. They didn't just ask random people to rewrite things; they hired expert translators who specialize in making things easy to read.

4. The Process: The "Human Craftsmen"

You might think, "Can't we just use AI to do this?" The paper says, "Not yet." AI is like a very smart student who has read a lot of books but hasn't learned the feeling of what makes something truly easy to understand.

So, the team acted like master craftsmen:

1. Selection: They picked real documents about voting, laws, and rights (the "ingredients").
2. Translation: Expert humans rewrote these documents into Easy-to-Read versions.
3. Annotation (The Secret Sauce): This is the most important part. As the experts rewrote the text, they kept a checklist. They wrote down exactly what they changed.
   • Did I break this long sentence into two? (Check)
   • Did I replace a hard word with a simple one? (Check)
   • Did I add a picture or a bullet point? (Check)

This checklist is like a "recipe card" attached to every dish. It tells future computers (and humans) exactly how the experts made the text easier.

5. The Result: A Gift to the World

The team has created a multilingual treasure chest (a corpus) containing thousands of these "Before and After" pairs in Spanish, Catalan, and Italian.

• Why Catalan is special: This is the first time anyone has ever made a high-quality collection of these texts for the Catalan language. It's like finding a lost map to a hidden island.
• Why it matters: Now, researchers can use this "treasure chest" to train AI. Eventually, computers will be able to look at a confusing government law and instantly turn it into a clear, friendly version for everyone.

The Big Picture Metaphor

Think of democracy as a giant, complex video game.

• The Problem: The instruction manual is written in a secret code that only a few people understand. Most players are stuck on the first level because they can't read the rules.
• The Paper's Contribution: The authors didn't just write a new manual; they built a translator machine. They gathered a team of experts to write down, step-by-step, how to turn the "secret code" into plain English (or Spanish/Catalan/Italian).
• The Future: Now, they are giving this translator machine to the whole world for free. Soon, anyone, regardless of their reading level, will be able to pick up the controller and play the game of democracy.

In short: This paper is about building a bridge so that no one is left out of the conversation about how we run our society. They built the bridge by hand, wrote down exactly how they did it, and are now handing the blueprints to everyone.

Technical Summary

Here is a detailed technical summary of the paper "A Multilingual Human Annotated Corpus of Original and Easy-to-Read Texts to Support Access to Democratic Participatory Processes."

1. Problem Statement

The paper addresses the critical lack of high-quality, parallel datasets for Automatic Text Simplification (ATS), specifically within Easy-to-Read (E2R) contexts for democratic participation.

• The Gap: While English has some resources (e.g., Simple English Wikipedia, Newsela), there is a severe scarcity of high-quality, human-annotated parallel corpora for Spanish, Catalan, and Italian. Existing resources often lack strict parallelism (only aligning comparable sentences), rely on non-expert crowd workers (leading to minimal simplification), or are restricted by copyright.
• The Context: Linguistic barriers prevent millions of people (including those with intellectual disabilities, low literacy, and migrants) from participating in democratic processes. Current Large Language Models (LLMs) require high-quality fine-tuning data to effectively perform text simplification, but such data is currently unavailable for these specific languages and domains.
• The Challenge: Creating a unified, high-quality corpus across three languages is difficult due to the lack of standardized E2R guidelines, differing linguistic needs (e.g., Catalan weak pronouns), and the tacit nature of expert simplification knowledge.

2. Methodology

The authors developed the iDEM Corpus through a rigorous, human-centric workflow involving the iDEM project consortium.

• Data Collection & Selection:
  • Domains: Texts were selected from political, deliberative, and participatory domains (government communication, news, legislative texts, social services).
  • Criteria: Selection prioritized relevance to democratic participation, copyright clearance, and ethical standards.
  • Institutions: 12 public, social, and private institutions collaborated to provide source texts.
• Annotation Schema & Simplification Rules:
  • A unified annotation schema was developed to map E2R criteria across the three languages, adapting to local standards (e.g., UNE for Spanish, DINCAT for Catalan).
  • Key Rules: The schema enforces specific linguistic transformations, including:
    • Lexical: Replacing technical jargon with frequent, simple words; defining complex terms.
    • Syntactic: Enforcing Subject-Verb-Object (SVO) order; breaking complex sentences into shorter segments; avoiding ellipsis and passive voice.
    • Formatting: Using line breaks to separate ideas (1 idea per line); using bullet points; avoiding unusual punctuation.
• Human Annotation Process:
  • Experts: Simplifications were performed by trained E2R professionals (3 for Spanish, 2 for Catalan, 2 for Italian) from reputable institutions, not crowd workers.
  • Workflow: Translators simplified texts sentence-by-sentence in spreadsheets. Crucially, they annotated which specific simplification rule they applied to each change.
  • Quality Assurance: Cross-checking was performed between translators to ensure consistency, though the high cost limited the use of multiple translators per text.
• Corpus Structure:
  • Format: JSON files containing metadata, original text chunks (~500 characters), simplified text, and alignment data.
  • Alignment: Sentence-level alignment (Original $\to$ Simplified). The corpus supports 1:1, 1:2, 1:3, and 1:N alignments (where one complex sentence is split into multiple simple segments).
  • Granularity: The unit of analysis is the "sentence segment" (a line of text containing one main idea), which may differ from grammatical sentences.

3. Key Contributions

• First-of-its-Kind Catalan Resource: This is the first available annotated corpus for sentence-level text simplification in Catalan.
• Substantial Expansion for Spanish and Italian: It provides a significant addition to the scarce data pools for Spanish and Italian, offering high-quality, expert-annotated parallel data.
• Rich Annotation: Unlike many existing datasets, this corpus includes fine-grained annotations indicating exactly which simplification criteria were applied (e.g., "reordering to SVO," "definition of technical term").
• Domain Specificity: The corpus focuses exclusively on democratic participation and deliberation, a domain where linguistic accessibility is critical but previously under-resourced.
• Open Access: The corpus is designed to be publicly available via a GitHub repository (scheduled for release in Summer 2026).

4. Results and Statistics

The paper provides detailed statistics on the corpus composition and simplification patterns:

• Scale:
  • Spanish: ~11,665 words (Original), ~10,883 words (Simplified), 354 original sentences $\to$ 1,290 simplified segments.
  • Catalan: ~11,570 words (Original), ~14,279 words (Simplified), 380 original sentences $\to$ 405 simplified segments.
  • Italian: ~10,398 words (Original), ~12,230 words (Simplified), 325 original sentences $\to$ 718 simplified segments.
• Simplification Intensity:
  • Sentence Segmentation: The simplification process significantly increased the number of segments. For example, in Spanish, the average words per segment dropped from 32.95 (original) to 8.44 (simplified).
  • Alignment Distribution:
    • 55.43% of original sentences were simplified 1:1.
    • 13.13% were split into 2 segments.
    • 7.55% were split into 5 or more segments (often due to complex lists or subordinate clauses).
• Most Frequent Simplification Criteria:
  • Spanish: Use of paragraph/line breaks (22.8%), frequent verbs (11.0%), and short nouns (10.5%).
  • Catalan: Choosing familiar words (55.3%) was the dominant strategy.
  • Italian: Use of common verbs (13.34%) and separating ideas with periods/commas (12.38%).

5. Significance and Impact

• Advancing NLP Research: The corpus provides the necessary "gold standard" data to train and evaluate generative AI and LLMs for text simplification in under-resourced languages.
• Democratizing Information: By providing resources specifically for political and civic discourse, the project directly supports the European Accessibility Act (EAA) 2025, which mandates accessible digital public content.
• Bridging the Expertise Gap: The paper highlights the difficulty of codifying "tacit knowledge" from human experts. The resulting annotation schema serves as a bridge between human E2R practices and machine learning requirements.
• Future Applications: The dataset is intended for use in the MER-TRANS IBERLEFT shared task and will serve as a foundation for future research in inclusive communication, cognitive accessibility, and multilingual NLP.

In summary, the iDEM Corpus represents a major step forward in making democratic processes linguistically accessible, filling a critical data void for Spanish, Catalan, and Italian through a rigorous, expert-driven, and transparent methodology.