OasisSimp: Multilingual Sentence Simplification

• OasisSimp is an open-source, human-annotated multilingual dataset for sentence simplification covering both high-resource (English) and low-resource languages.
• It integrates diverse corpora and rigorous manual curation to generate parallel complex–simple sentence pairs across English, Sinhala, Tamil, Pashto, and Thai.
• The benchmark employs zero-shot and few-shot prompting with metrics like SARI and BERTScore to reveal LLM performance discrepancies across languages.

Searching arXiv for OasisSimp and related sentence simplification papers. OasisSimp is an open-source, human-annotated multilingual dataset for sentence-level simplification covering English, Sinhala, Tamil, Pashto, and Thai. It was introduced to address the scarcity of high-quality parallel complex–simple sentence pairs for mid-resource and low-resource languages, especially beyond English, and is positioned as both a resource and a benchmark for studying multilingual simplification under zero-shot and few-shot prompting regimes (Liu et al., 14 Mar 2026). The name should be distinguished from other datasets built around the acronym “OASIS,” including a culturally grounded multimodal QA resource (Alam et al., 7 Oct 2025), a water-sustainability dataset (Gupta et al., 2024), and a single-image 3D dataset (Chen et al., 2020).

1. Scope, motivation, and linguistic coverage

OasisSimp was designed around a specific gap in sentence simplification research: while simplification has matured for high-resource languages, especially English, progress remains limited for mid-resource and low-resource languages because high-quality parallel data is scarce (Liu et al., 14 Mar 2026). The dataset therefore targets five languages with sharply different resource conditions. English functions as the high-resource comparison language, whereas Sinhala, Tamil, Pashto, and Thai were selected because they are comparatively low-resource in simplification and present distinct scripts, typological properties, and sentence-formation challenges.

The language selection is central to the dataset’s role in the literature. The paper states that no prior sentence simplification datasets exist for Thai, Pashto, and Tamil, while Sinhala had only a small existing resource. This gives OasisSimp a dual function: it supports direct benchmark evaluation in underrepresented languages and provides a controlled setting for studying multilingual transfer, prompting, and generalization when the simplifying language is not strongly represented in existing resources (Liu et al., 14 Mar 2026).

The broader motivation is also practical and social. The dataset is framed as relevant to literacy, education, accessibility, and information access for languages spoken by large populations. At the same time, its technical emphasis remains sentence-level simplification rather than document adaptation, readability scoring, or lexical substitution in isolation.

2. Source corpora and data selection

OasisSimp was constructed through a manually controlled pipeline rather than automatic alignment or synthetic generation. Because there was no common source across all languages, the authors selected different public corpora per language and then filtered for sentence-level complexity and diversity (Liu et al., 14 Mar 2026).

For English, sentences came from Canadian newspaper text via the NewsEdits workflow, primarily The Globe and Mail. The selection process imposed restrictions such as 100–300 characters, removal of templated material, and limiting proper nouns to reduce reliance on world knowledge. This yielded 2,500 context-independent complex English sentences.

Sinhala and Tamil were drawn from the SiTa trilingual parallel corpus of government documents. For Sinhala, rare-word filtering was applied using Common Crawl-based frequency counts, after which the team manually removed list-like or near-duplicate sentences before selecting 2,500 long, complex sentences. Some bracketed English explanations were removed. Tamil was sampled from the Sinhala-selected set, but annotation capacity limited the final Tamil dataset to 520 sentences.

Thai sentences came from ThaiSum news data. The authors used a Thai sentence segmenter, then filtered for length, rare and non-rare words, and topic diversity across ten news categories; sentence boundaries were manually corrected to produce 1,499 final sentences. Pashto sentences came from Wikipedia, with 10,000 candidates initially sampled evenly across ten semantic categories; human annotators then selected the most complex 250 per category, yielding 2,500 sentences. Native linguistic review was used to ensure cultural and linguistic appropriateness (Liu et al., 14 Mar 2026).

This construction strategy suggests that OasisSimp is not a parallel corpus derived from a uniform multilingual source. A plausible implication is that cross-language comparisons should be interpreted as comparisons of simplification difficulty under heterogeneous source domains as well as heterogeneous language conditions.

3. Annotation protocol and operational definition of simplification

The dataset was annotated by 3–6 native speakers per language, all with at least a bachelor’s degree. Annotators were recruited through the authors’ own contacts in the relevant countries rather than through an open online platform, with the stated goal of improving quality control. They received structured training and at least three training rounds, and annotation proceeded in batches of 25–100 sentences depending on language and annotator capacity (Liu et al., 14 Mar 2026).

The guidelines were adapted from prior simplification work and required annotators to simplify while preserving meaning, fluency, and grammaticality. Four operations were explicitly taught and illustrated: rewording difficult or technical terms with simpler synonyms, sentence splitting, deletion of unnecessary details, and reordering for clarity. The paper defines the task as sentence-level syntactic simplification: transforming a complex sentence into a simpler sentence while preserving the original meaning and factual content.

Quality control relied on iterative supervision rather than a formal inter-annotator agreement statistic. The authors initially checked the work closely, corrected misunderstandings through further instruction, and later switched to random sampling for verification once annotator quality had been established. The paper also notes that some English sentences were excluded if quality issues arose. This means that dataset quality is grounded in guided human supervision and spot checking rather than agreement-based validation (Liu et al., 14 Mar 2026).

A common misconception is to treat sentence simplification as primarily lexical replacement. OasisSimp explicitly does not do that. Its operationalization includes deletion, splitting, and reordering in addition to lexical rewording, and the benchmark is framed as sentence simplification rather than lexical-only simplification.

4. Dataset composition and split design

In aggregate, OasisSimp contains 9,019 complex sentences, with multiple simplifications per source sentence (Liu et al., 14 Mar 2026). The dataset is split into validation and test subsets, with 80% of the data used for testing and 20% for validation. The paper presents this split as appropriate for unsupervised or prompt-based development rather than supervised training, because the dataset is still too small to train large simplification models from scratch.

Language	Complex sentences	Simplified sentences per source
English	2,500	2.86
Sinhala	2,500	5.00
Thai	1,499	5.06
Tamil	520	4.66
Pashto	2,500	3.00

The per-language averages further characterize the corpus. English has an average complex length of 24.35 words and an average simplified length of 17.23 words, sourced from news. Sinhala has 30.12 and 28.78, sourced from government documents. Thai has 48.24 tokens and 37.77 tokens, sourced from news. Tamil has 23.22 and 17.65, sourced from government documents. Pashto has 28.81 and 20.31, sourced from Wikipedia (Liu et al., 14 Mar 2026).

These statistics indicate that simplification in OasisSimp is not uniformly realized as aggressive shortening. Sinhala, for example, shows a relatively small average reduction in length, which suggests that simplification may often be achieved through rephrasing and structural clarification rather than extensive content deletion.

5. Evaluation framework and benchmarked models

The paper evaluates eight open-weight multilingual LLMs using the prompting strategy of BLESS, instructing the model to “Simplify the given sentence …” (Liu et al., 14 Mar 2026). The evaluated models are Aya (Aya-Expanse-8B), Cmd-R (c4ai-command-r7b-12-2024), DeepSeek (deepseek-LLM-7B-chat), EuroLLM (EuroLLM-9B-Instruct), Gemma (Gemma-3-12B-it), LLaMA (Llama-3.2-3B-Instruct), Mistral (Mistral-7B-Instruct-v0.2), and Qwen (Qwen2.5-7B-Instruct). Evaluation is reported in zero-shot, 1-shot, and 5-shot settings.

For zero-shot prompting, temperatures from 0.1 to 0.9 in steps of 0.1 were searched and the best result was reported. Few-shot experiments used the best zero-shot temperature. The main automatic metrics were SARI and BERTScore (F$_\text{ref}$). BLEU was reported in the appendix for completeness, but the paper explicitly did not rely on it because BLEU is known to be poorly suited for simplification evaluation. SARI was further broken down into add, keep, and delete sub-scores.

The benchmark is therefore designed less as a supervised learning leaderboard than as a diagnostic evaluation of multilingual prompting behavior. This suggests that OasisSimp is particularly suited to probing how contemporary multilingual LLMs handle simplification when language coverage is uneven and direct finetuning data is limited.

6. Empirical findings, limitations, and research significance

The central empirical finding is that model performance is substantially better on high-resource English than on the low-resource languages, and that few-shot prompting consistently helps (Liu et al., 14 Mar 2026). In English, scores vary relatively little across models, which the paper associates with stronger language coverage in pretraining and the availability of abundant simplification data. In contrast, Pashto and Thai show the largest variability across models in zero-shot settings, indicating that off-the-shelf multilingual LLMs do not reliably generalize to low-resource simplification.

Gemma is identified as the most robust and consistent model across languages, with especially strong BERTScore on Sinhala, Tamil, and Pashto. DeepSeek and LLaMA show particularly poor behavior in some low-resource settings, including negative or near-zero BERTScore values in certain cases, which the paper interprets as severe failure to match the reference simplifications semantically. Few-shot prompting improves performance across nearly all languages and models. The paper gives concrete examples for Gemma, whose SARI rises from 39.4 to 43.6 on English, 34.8 to 39.9 on Sinhala, 37.9 to 41.3 on Thai, 32.8 to 39.3 on Tamil, and 33.3 to 37.9 on Pashto.

At the component level, deletion is the easiest operation for LLMs, while addition is the hardest. Models are described as good at removing redundant information but weak at generating new, simpler phrasing, especially in low-resource and morphologically complex languages. KEEP scores improve in few-shot settings, suggesting that in-context examples help models preserve relevant content, but ADD remains consistently low (Liu et al., 14 Mar 2026).

The paper also states several limitations. Language selection depended on available annotators, so the benchmark is not a systematic typological study. The dataset is relatively small and constrained by annotation budget and available expertise, which prevented training large supervised simplification systems. Human evaluation of LLM outputs was not conducted, so the benchmark relies on automatic metrics. Because the pretraining corpora of the evaluated LLMs are not fully transparent, strong causal claims about the effect of language inclusion in pretraining are not possible. Deeper study of cross-linguistic transfer and language-family relationships is also beyond scope.

Ethically, the source sentences come from public, authentic materials such as government documents, news, and Wikipedia. Annotators were instructed to preserve meaning, which the authors present as reducing the risk of generating undesirable content. Annotators were paid at local standard rates or offered co-authorship. One caveat explicitly acknowledged is that not every sentence was individually inspected for undesirable content (Liu et al., 14 Mar 2026).

Taken together, OasisSimp functions as a benchmark for zero-shot and few-shot multilingual sentence simplification, especially in underrepresented languages. Because it includes multiple simplifications per source sentence, it is also relevant for studying metric robustness and reference sensitivity. More broadly, it provides a human-created basis for work on multilingual transfer, controllable generation, and language-specific simplification strategies in settings where linguistic resources are scarce.