Public DGS Corpus: Multilingual E2R Simplification

• Public DGS Corpus is a multilingual resource pairing original texts with expert-produced Easy-to-Read simplifications in Spanish, Catalan, and Italian.
• It supports analysis of accessibility in public communication by providing detailed alignment, segmentation, and annotated simplification criteria.
• Its methodology leverages human expertise and rigorous quality control to benchmark automatic text simplification and civic discourse studies.

Searching arXiv for the most relevant paper on the Public DGS Corpus / iDEM corpus. arxiv_search(query="Public DGS Corpus iDEM Easy-to-Read democratic participation corpus", max_results=5) The Public DGS Corpus, described in the paper as the iDEM Corpus, is a multilingual, human-annotated corpus of original and Easy-to-Read (E2R) simplified texts designed to support research and applications in democratic participation, deliberation, and accessible public communication. It was developed within the Horizon Europe iDEM project to assess the impact of Easy-to-Read language for democratic participatory processes, and it pairs source texts with expert-produced simplifications and annotation of the simplification criteria applied. The resource covers Spanish, Catalan, and Italian, is aligned at the document and sentence level within each language, and is intended for training and evaluation in automatic text simplification as well as for broader accessibility research in public-domain, democracy-related settings (Bott et al., 5 Mar 2026).

1. Terminology, scope, and conceptual orientation

The paper uses the name iDEM Corpus in its main text, while the designation Public DGS Corpus reflects the framing of the resource as a public corpus for democracy-related accessibility research. In substance, the resource is defined by three properties: it is multilingual, human-annotated, and centered on original/E2R simplification pairs in domains connected to democratic participation (Bott et al., 5 Mar 2026).

Its stated rationale is explicitly political as well as computational. The corpus is motivated by the premise that access to information is a prerequisite for participation in society, and especially in democratic deliberation. The paper situates people with reading difficulties, people with disabilities, migrants, and other groups as populations that may be effectively excluded from civic life when institutional or political texts are linguistically too complex. The corpus therefore functions both as a language-technology resource and as an infrastructure for accessibility-oriented public communication.

The resource is also positioned against a scarcity of high-quality simplification data. The paper states that such resources are limited for English and even more limited for Spanish, Catalan, and Italian, with the Catalan component described as the first annotated corpus of its kind for the Catalan language. This gives the corpus a dual significance: it is a domain-specific democracy corpus and a multilingual simplification corpus for comparatively less resourced settings.

2. Multilingual composition and domain coverage

The corpus contains three language-specific subcorpora: Spanish, Catalan, and Italian. These are not parallel across languages in the sense of cross-lingual translations of the same source documents. Instead, each language subcorpus was created independently under a shared methodology and common JSON structure. The paper is explicit that the corpus is parallel on the document/sentence level within each language, because each original text is paired with its simplified version, but not parallel across languages (Bott et al., 5 Mar 2026).

The original texts were selected from domains directly related to democratic participation and public access to information. The corpus includes material from:

• government communication
• news
• political discourse
• legislative and policy texts
• public/social service documents
• informative texts
• advocacy and civic participation materials

The topical range includes civic participation, justice, elections, EU institutions, social policy, data protection, health governance, legislation, environmental policy, disability accessibility, children’s rights, political activism, and deliberative democracy. This thematic profile is central to the corpus definition: the texts were chosen not as generic simplification material, but as language that conditions meaningful access to public and democratic processes.

Text selection followed three stated criteria: relevance to the iDEM project and democratic participation, copyright availability / clearance, and ethical standards. The paper further states that selection was done collaboratively within the consortium through internal voting, and that 12 public, social, and private institutions participated in creating the corpus.

3. Construction methodology and annotation principles

A defining feature of the corpus is that simplification was carried out by trained human experts in Easy-to-Read, not by automatic systems and not by non-specialist crowd workers. The simplifiers also annotated the simplification criteria they applied, so the annotation scheme served simultaneously as a production guideline and as a linguistic annotation layer bundled with the final corpus (Bott et al., 5 Mar 2026).

The methodology is grounded in E2R standards and local adaptations. The paper states that the annotation schema was influenced by:

• IFLA guidelines
• Inclusion Europe
• ISO/IEC 23859:2023
• ISO 24495-1:2023
• the UNE standard for Spanish
• DINCAT resources for Catalan

This standardization effort is presented as necessary because E2R is not a protected designation and lacks universally enforced standards. A plausible implication is that the corpus is not merely a data collection exercise, but also an attempt to operationalize and harmonize heterogeneous simplification traditions across languages.

Across the three subcorpora, the simplification guidelines emphasize the use of frequent, common words; avoidance of abstract terms, foreignisms, vague content, and technical jargon; preference for short, simple nouns and common verbs, adjectives, and adverbs; explicit definition of difficult terms when necessary; simple syntax; preference for subject–verb–object order; splitting ideas with paragraph breaks, line breaks, periods, and commas; avoidance of uncommon punctuation and complex formatting; and expansion of acronyms and abbreviations when needed.

The paper also records language-specific guidelines. For Catalan, a special rule prohibits using more than one weak pronoun together. For Spanish, the guidelines emphasize simple tenses and explicit addition of subjects to avoid ellipsis. For Italian and Spanish, line-breaking is treated as a core E2R device. The examples in the paper show that simplification is not restricted to lexical substitution. It also includes syntactic restructuring, discourse segmentation, formatting changes, and explicit contextualization, such as explaining the acronym ONU, replacing “rectificar” with “canviar o corregir,” or replacing “aleatoriamente” with “al azar.”

4. Corpus size, alignment patterns, and quantitative profile

The corpus statistics are reported separately for each language. The following table summarizes the core sentence- and word-level counts.

Language	Sentence counts	Word counts
Spanish	354 original; 1290 simplified sentence segments	11,665 original; 10,883 simplified
Catalan	380 original; 405 simplified sentence segments	11,570 original; 14,279 simplified
Italian	325 original; 718 simplified sentence segments	10,398 original; 12,230 simplified

The paper also reports average words per sentence segment:

• Spanish: 32.95 original vs. 8.44 simplified
• Catalan: 41.32 original vs. 35.26 simplified
• Italian: 31.99 original vs. 17.03 simplified

These figures show that simplification often increased segmentation granularity rather than merely shortening texts. The paper explicitly interprets the Spanish statistics as evidence that Spanish translators followed the E2R recommendation to place each main idea on a separate line much more strictly than Catalan translators did (Bott et al., 5 Mar 2026).

Alignment statistics further characterize the rewriting process:

• 1:1 = 587 alignments (55.43%)
• 1:2 = 139 (13.13%)
• 1:3 = 116 (10.95%)
• 1:4 = 88 (8.31%)
• 1:5 = 49 (4.63%)
• 1:>5 = 80 (7.55%)

The paper notes that the 1:>5 cases often reflect list-like structures or highly complex originals with multiple subordinate and coordinate clauses. It also states that the workflow did not allow many-to-one or many-to-many relations, nor deletions represented as 1:0 alignments; such cases were excluded by design. This restriction means that the corpus operationalizes simplification primarily as expansion and segmentation of original units, not as unrestricted restructuring.

The corpus also reports category distributions by language. For Spanish, the original-word distribution is 4892 informative texts (41.94%), 1146 political/ideological articles (9.82%), 3817 news articles (32.72%), and 1810 policy/legislative documents (15.52%). For Catalan, it is 6090 informative texts (52.64%), 1652 social justice/public policy analysis (14.28%), and 3828 policy/administrative documents (33.09%). For Italian, it is 4425 informative texts (42.56%), 1247 institutional/legal documents (11.99%), 2680 political/legal analysis (28.24%), and 1789 advocacy/social responsibility (17.21%).

5. Data model, preprocessing, and quality control

The corpus is released in JSON format, with one file per language and integrated metadata. The structure includes document-level metadata such as document_id, language code, source URL, character counts, corpus category, text type, topic, simplification level, methodology, copyright status, and translator and institution information. It also contains text chunks of about 500 characters without spaces, original and simplified sentences as separate objects with IDs, sentence-level alignments, and annotations on each simplified sentence indicating complex components and applied simplification criteria (Bott et al., 5 Mar 2026).

The paper states that the text_id enables tracing simplification at the sentence, text, and document levels. This hierarchical design is methodologically important because it supports multi-level analysis of rewriting operations rather than limiting study to flat sentence pairs.

Quality control relied on expert participation and partial cross-checking rather than full redundancy. The paper states that two project partners brought decades of expertise and teams of trained language professionals; the corpus was created by experienced institutional collaborators; and there were cross-checks among annotators and translators. The translator distribution was three for Spanish, two for Catalan, and two for Italian. Because of budget constraints, however, not every item received full double annotation. Instead, random text units were sampled and cross-checked between translators.

This quality-control design has clear implications. On one hand, expert production is likely to yield higher-fidelity E2R rewriting than non-expert annotation. On the other, the absence of full multi-annotator redundancy means that inter-annotator variation cannot be exhaustively characterized from the corpus construction protocol itself. The paper acknowledges this as a limitation rather than presenting the annotations as fully standardized in a narrow reproducibility sense.

6. Access conditions, limitations, and relation to adjacent resources

The paper states that the corpus will be made available through a public GitHub repository, scheduled for release in summer 2026, and that release was delayed because the data will be used in the MER-TRANS IBERLEFT shared task. It also states that the corpus is intended to be freely accessible to the public. At the same time, not all associated materials are equally open: the Spanish and Italian simplification guidelines are the intellectual property of Capito, whereas the Catalan guidelines and the annotation tagsets will be released with the corpus (Bott et al., 5 Mar 2026).

The paper is candid about several limitations. It notes the absence of universal official E2R guidelines across languages; the inevitability of bias and error in human annotation; the possibility that experts may not be accustomed to explicitly annotating their own tacit simplification decisions; and constraints imposed by copyright and ethical exclusions. These limitations do not negate the corpus’s utility, but they do define its methodological status: it is a high-quality expert corpus produced under real-world legal, budgetary, and standardization constraints.

A recurrent source of ambiguity concerns the label DGS itself. In adjacent literatures, DGS may evoke unrelated resources, including corpora tied to other languages, modalities, or domains. The resource discussed here is instead a multilingual E2R text corpus for democratic participation, centered on Spanish, Catalan, and Italian. In the broader landscape of democracy-oriented corpora, it differs substantially from resources such as GDN-CC, which focuses on French citizen-consultation data, argumentative-unit extraction, and clarification of contributions from the Grand Débat National rather than expert E2R simplification of public-information texts (Lequeu et al., 21 Jan 2026). This suggests that the Public DGS / iDEM corpus occupies a distinct niche at the intersection of text simplification, accessibility, and democratic communication.

Its broader significance follows directly from that niche. The corpus links language accessibility to democratic inclusion, provides expert-produced aligned simplifications in politically relevant domains, and supplies a reusable benchmark for research on E2R rewriting in three languages. It is therefore both a domain corpus for public communication and a methodological resource for studying how institutional and civic texts can be transformed into more accessible forms without departing from the informational requirements of democratic participation.