SAMER Corpus: Arabic Text Simplification Dataset

• SAMER Corpus is a manually annotated, three-way parallel dataset featuring original texts with two simplified versions aimed at school-aged learners.
• It uses a staged lexical simplification process with exact alignment at document, paragraph, sentence, and fragment levels for rigorous readability control.
• The resource supports research in Arabic automatic readability assessment and pedagogical language technologies with detailed transformation statistics.

Searching arXiv for the SAMER corpus paper and closely related Arabic simplification/readability resources. The SAMER Arabic Text Simplification Corpus is a publicly available Arabic parallel corpus for text simplification targeting school-aged learners. It is presented as the first manually annotated Arabic parallel corpus for text simplification, and it combines original literary texts with two simplified versions, document-level readability labels, and word-level readability labels. The resource was designed to support Arabic text simplification, Arabic automatic readability assessment, and the development of pedagogical language technologies, with a simplification regime defined as paraphrasing into a controlled language anchored in a readability-leveled lexicon (Alhafni et al., 2024).

1. Scope, provenance, and intended use

SAMER was created within the broader SAMER project, expanded in the paper as “Simplification of Arabic Masterpieces for Extensive Reading.” The project also includes a readability-leveled Arabic lexicon, an online readability thesaurus, and a Google Docs add-on for Arabic readability visualization and assisted simplification. Within that larger program, the corpus functions as a manually revised benchmark and as an educational resource.

The corpus contains about 159K words drawn from 15 publicly available Arabic fiction novels. The source selection was constrained by four stated criteria: Arabic fiction novels, large historical span, high readability / difficult texts, and publicly available status, either because the texts were out of copyright or under an open license. Budget constraints limited the collection to 15 books.

The selected works are historically heterogeneous but not genre-diverse in the broad corpus sense, since they are all fiction novels. The paper states that 14 of the novels were published between 1865 and 1955, and that one philosophical novel dates to the 12th century. The 15 books are:

1. حي بن يقظان (Hayy Ibn Yaqzan) — Ibn Tufail — 12th century philosophical novel
2. غابة الحق (The Forest of Truth) — Francis Marrash — 1865
3. لادياس (Ladiyas) — Ahmed Shawqi
4. المحالفة الثلاثية في المملكة الحيوانية (The Tripartite Alliance of the Animal Kingdom) — Ameen Rihani
5. الملك كورش (Cyrus the Great) — Zaynab Fawwaz
6. الأجنحة المتكسرة (Broken Wings) — Kahlil Gibran
7. زينب (Zaynab) — Mohammed Hussein Heikal
8. شجرة الدر (The Pearl Tree) — Jurji Zaydan
9. إبراهيم الكاتب (Ibrahim Al-Katib) — Ibrahim Abd Al-Qadir Al-Mazini 10. ثورة في جهنم (A Revolution in Hell) — Niqula Haddad
10. سارة (Sara) — Abbas Mahmoud Al-Aqqad
11. فارس بني حمدان (The Knight of Beni Hamdan) — Ali Al-Jarem
12. على باب زويلة (On Bab Ziwaila) — Mohammed Saeed Al-Aryan
13. نماذج بشرية (Human Examples) — Ahmad Rida Huhu
14. هذا التاج (This Crown) — Wasef Al-Baroudi

The educational target is explicit. SAMER provides two simplified outputs aimed at two grade bands: Level 4 (L4) for approximately grades 6–8, ages 11–14, and Level 3 (L3) for approximately grades 4–5, ages 9–10. This positions the corpus not merely as an NLP dataset but as a resource for level-appropriate reading materials.

2. Parallel architecture and alignment design

SAMER is a three-way parallel corpus. Each document appears in three aligned forms: Original text, Level 4 simplified text, and Level 3 simplified text. The simplification is staged rather than jointly produced. The paper defines the stages as follows:

• Original → L4: all words above L4, especially Level 5 items, are simplified to Level 4 or lower
• L4 → L3: all remaining Level 4 words are simplified to Level 3 or lower

This design makes the resource suitable for controlled simplification experiments in which one can model either the first simplification step, the second step, or the full original-to-L3 transformation.

A central structural property is exact alignment. The paper states that “all three versions of each document will have the same number of paragraphs and sentences by design.” Later, the texts were segmented into punctuated fragments, and because punctuation was preserved during annotation, those fragments are described as perfectly aligned across versions. This gives the corpus aligned supervision at several granularities: document, paragraph, sentence, and punctuation-delimited fragment.

The main size figures for the three document streams are as follows:

Version	Words	Stated readability constraint
Original	159,265	Original documents are effectively Level 5
Level 4 simplified	159,677	All level 5 words simplified to level 4 or lower
Level 3 simplified	160,149	All level 4 words in the Level 4 text simplified to level 3 or lower

The initial extraction process took roughly the first 10K words from each novel while respecting chapter boundaries. This yielded 4,289 paragraphs, which were later segmented into 20,603 punctuated fragments. The average fragment length is about 7.5 words.

3. Readability model and annotation scheme

The corpus is organized around a readability model derived from the SAMER readability lexicon. The lexicon uses a five-level scale:

• L1: easiest
• L2
• L3
• L4
• L5: hardest

The paper gives representative examples for these levels:

• L1: بَيْت، شَجَرَة، صَنَعَ، لكِن
• L2: جَزيرة، داكِن، خَدَعَ، إذا
• L3: مُتْحَف، رِئة، لدى، لِكَيْ
• L4: اِقْتِصاد، طُمَأنينة، راقِ، نَكَثَ
• L5: قَسْطَرة، هَيْضة، لَوْذَع، شُعَبِيّ

For annotation, the operational label set is broader than the pedagogical five-level scale. The corpus uses Level 0 for proper nouns and Level 6 for unknown words not found in the lexicon. The practical word-level labels are therefore L0–L6.

The document-level rule is also explicit and unusually strict:

“For the purpose of the corpus annotation, we consider the document readability level to be equal to the highest readability level found among the words in the document.”

Under that definition, the selected source texts are effectively all Level 5 documents, because they were intentionally chosen as difficult literary works aimed originally at proficient Arabic readers. The simplification objective is therefore not to average difficulty downward, but to guarantee that the maximum lexical difficulty in a document does not exceed the target threshold.

This lexicon- and lemma-driven design reflects a core linguistic motivation in the paper. Arabic is described as morphologically rich and orthographically ambiguous, and the authors emphasize that lexical simplification cannot be reduced to surface-form substitution because one orthographic form may correspond to multiple lemmas and readings. A plausible implication is that SAMER was designed to support simplification systems that operate on disambiguated lexical representations rather than on raw token strings.

4. Annotation workflow, simplification policy, and validation

The annotation process combined automatic pre-annotation and manual expert revision. The work was carried out by three professional female computational linguists, all native Arabic speakers, through a linguistic annotation firm. Annotation was performed in 146 Google Docs, each equipped with the SAMER Google Docs add-on (Alhafni et al., 2024).

The workflow proceeded in seven stated steps:

1. Texts were automatically tokenized and morphologically disambiguated.
2. The system used the BERT unfactored morphological disambiguator of Inoue et al. (2022) through CAMeL Tools.
3. For each word, the predicted lemma and POS tag were looked up in the readability lexicon of Al Khalil et al. (2020).
4. The add-on highlighted words in different colors by readability level.
5. Annotators corrected wrong levels, especially errors caused by lexicon coverage gaps and morphological disambiguation errors.
6. Annotators simplified the text while preserving meaning and grammaticality.
7. They reran the readability analysis to verify the target document level had been reached.

The simplification regime is deliberately narrow. The paper states that the authors focus on lexical simplification, not syntactic rewriting. Their rationale is twofold: lexical simplification is known to improve readability, and it allows semantic equivalence to be preserved while avoiding the complications of restructuring Arabic syntax. The permitted operations are replacements, deletions, and insertions. These operations may involve more than one word, but the original meaning must be preserved, the output must remain grammatical, and no abridgement or summarization is allowed.

From this policy, the paper derives a practical taxonomy of transformations: one-to-one replacements (1-1), one-to-many replacements (1-m), many-to-one replacements (m-1), insertions, and deletions. It explicitly notes that many-to-many (m-m) replacements do not occur in the corpus. This is significant because it indicates a tightly constrained editing protocol rather than unconstrained rewriting.

Quality control combines process constraints with double annotation. Annotators had to verify and correct word-level readability anomalies, rerun the add-on after revisions, and confirm that the target readability level had been reached. For validation, the authors selected about 1,300 words from each book, totaling 17 paragraphs or roughly 20K words, for double annotation. The analysis is reported through mismatch rates rather than a kappa coefficient. The paper gives ~6.8% word-level mismatches for Original → L4 and 13.2% mismatches for L4 → L3. It further states that most mismatches reflected different lexical simplification choices, not annotation errors, and that actual annotation mistakes were ~10% of all mismatches. This means the validation emphasizes controlled human review rather than a single agreement scalar.

5. Quantitative profile

The quantitative profile of SAMER is defined both by readability distributions and by transformation statistics. For the original 159,265 words, the word-level distribution is:

• L0: 2,631 (1.7%)
• L1: 83,772 (52.6%)
• L2: 23,103 (14.5%)
• L3: 22,517 (14.1%)
• L4: 14,965 (9.4%)
• L5: 9,463 (5.9%)
• L6: 2,814 (1.8%)

For the L4 simplified version with 159,677 words, the distribution becomes:

• L0: 5,212 (3.3%)
• L1: 90,232 (56.5%)
• L2: 26,297 (16.5%)
• L3: 24,630 (15.4%)
• L4: 13,306 (8.3%)
• L5: 0
• L6: 0

For the L3 simplified version with 160,149 words, the distribution is:

• L0: 5,246 (3.3%)
• L1: 95,898 (59.9%)
• L2: 30,015 (18.7%)
• L3: 28,990 (18.1%)
• L4: 0
• L5: 0
• L6: 0

The authors summarize these shifts as follows: Original → L4 produces an 8.8% overall shift from higher to lower readability levels; L4 → L3 produces an 8.3% change; and Original → L3 yields a 17.1% decrease in overall readability levels (Alhafni et al., 2024).

Transformation statistics were derived with a word alignment tool based on character edit distance. For Original → L4, the paper reports:

• No Change: 152,214 (95.5%)
• 1-1 replacements: 6,430 (4.0%)
• 1-m replacements: 0.2%
• m-1 replacements: 0.1%
• Insertions: 0.1%
• Deletions: 0.0%

For L4 → L3, it reports:

• No Change: 145,090 (90.8%)
• 1-1 replacements: 13,508 (8.5%)
• 1-m replacements: 0.3%
• m-1 replacements: 0.1%
• Insertions: 0.2%
• Deletions: 0.1%

These figures show that the corpus is dominated by lexical substitution, especially one-to-one replacement, with only very limited insertion and deletion. That distribution is consistent with the explicitly lexical orientation of the annotation guidelines.

At the fragment level, the 20,603 punctuated fragments are divided into four categories:

• No Change: 8,920 (43.3%)
• Change in L4 only: 2,610 (12.7%)
• Change in L3 only: 6,369 (30.9%)
• Change in both L4 and L3: 2,704 (13.1%)

For reproducible experimentation, the paper provides chapter-based train/dev/test splits, keeping chapter integrity and selecting development and test chapters from well-separated regions of each novel to reduce topical leakage. The final split sizes are:

• Train: 113,476 words (71%)
• Dev: 22,280 words (14%)
• Test: 23,509 words (15%)

6. Research significance, limitations, and relation to the wider Arabic corpus landscape

SAMER is positioned as a corpus for Arabic text simplification, Arabic automatic readability assessment, pedagogical language technologies, and potentially personalized Arabic pedagogical applications. The paper does not report experimental baselines, model benchmark scores, or task leaderboards; its contribution is the construction, documentation, and release of the corpus rather than the evaluation of specific models (Alhafni et al., 2024).

The release details are practical rather than formalized as a license specification. The paper states that the corpus, its parallel versions, its splits, and the annotation guidelines are publicly available on the SAMER Project website: http://samer.camel-lab.com/. It does not state a named software or data license such as CC BY or GPL in the text provided. It does state that the source texts are in the public domain and that the corpus itself is publicly accessible through the project website.

Several common misunderstandings are precluded by the paper’s design choices. SAMER is not a corpus of unconstrained rewriting; it is a corpus of controlled lexical simplification. It is not presented as a syntax-focused simplification resource. It is not primarily a historical Arabic corpus, even though its source novels have a historical span extending from the 12th century through texts published between 1865 and 1955. It is also not a benchmark paper with system comparisons.

Its position becomes clearer when contrasted with Shamela, a large-scale historical Arabic corpus of about 1 billion words and more than 6,100 texts, built from Al-Maktaba Al-Shamela and oriented toward diachronic analysis, morphological processing, parallel-passage detection, and automatic dating (Belinkov et al., 2016). SAMER, by contrast, contains about 159K words from 15 fiction novels and is designed around manual readability annotation and parallel simplification. This suggests that the two resources occupy different methodological niches: Shamela is a resource for large-scale historical and intertextual study, whereas SAMER is a resource for controlled educational simplification and readability-sensitive supervision.

Within Arabic NLP, SAMER’s distinct contribution lies in combining several properties that are often separated across resources: manual revision, explicit readability targeting, two-step simplification, exact alignment across versions, and public availability. For researchers, its main value is therefore not scale but annotation design: the corpus makes it possible to study how lexical difficulty constraints propagate from word labels to document labels and how controlled transformations can be learned from aligned Arabic literary text.