RALS: Resources and Baselines for Romanian Automatic Lexical Simplification

2026-07-23 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, extended

Summary

The RALS project introduces the first comprehensive resources for Romanian lexical complexity prediction (LCP) and lexical simplification (LS), addressing a significant gap for under-represented languages. It provides a new dataset with 3,921 human-annotated word samples for LCP, comprising three subsets: HT (human-translated English MultiLS), WT (words from a Romanian corpus), and RoLCP (new samples from diverse Romanian texts). A novel RoLS dataset for lexical substitution candidates is also created, utilizing a robust pairwise simplicity ranking methodology. The paper evaluates LCP models using a Ridge regressor with handcrafted features, achieving a Pearson r=0.71 on the RoLCP dataset. For LS, it compares open LLMs (Apertus-8B-Instruct-2509, RoLlama3.1-8B) and a fine-tuned BERT with a new hybrid rule-based system called DexFlex. DexFlex, which integrates the dexonline dictionary and spaCy for grammatical processing and synonym inflection, consistently outperforms other approaches, though overall scores remain lower than for high-resource languages.

Key takeaway

For NLP Engineers developing text simplification systems for low-resource languages like Romanian, you should prioritize hybrid rule-based approaches over large language model prompting. The DexFlex framework demonstrates that combining dictionary data with grammatical processing can achieve state-of-the-art lexical simplification, outperforming LLMs. This approach is more sustainable and linguistically grounded, offering a robust alternative to costly, less transparent proprietary systems. Consider integrating high-quality linguistic resources into your pipelines.

Key insights

RALS provides the first Romanian lexical simplification resources, with a hybrid rule-based system outperforming LLMs for this low-resource language.

Principles

• Cross-lingual complexity transfer causes distribution shifts.
• Hybrid rule-based systems can outperform LLMs for LS.
• Simpler LCP models with handcrafted features are viable.

Method

DexFlex uses spaCy for POS/grammatical features, selects synonyms from dexonline via contextual embedding similarity, and inflects them correctly using dictionary data.

In practice

• Develop simpler LCP baselines with handcrafted features.
• Integrate dictionaries into NLP pipelines for low-resource LS.
• Use pairwise ranking for robust simplicity judgments.

Topics

• Lexical Simplification
• Romanian NLP
• Lexical Complexity Prediction
• Low-Resource Languages
• DexFlex Framework
• Natural Language Processing

Code references

• senisioi/RALS
• petruTH/DexFlex

Best for: AI Scientist, NLP Engineer, Research Scientist

Related on AIssential

• RALS: Resources and Baselines for Romanian Automatic Lexical Simplification — RALS provides the first integrated dataset and system for Romanian lexical simplification and complexity prediction.
• The Tatoxa System for Text Detoxification in Low-Resource Languages: The Case of Tatar — The Tatoxa system offers a superior solution for Tatar text detoxification, outperforming LLMs and highlighting native data's importance.
• A Calculus-Based Framework for Determining Vocabulary Size in End-to-End ASR — A calculus-based framework formally determines optimal vocabulary size for end-to-end ASR systems.
• Evaluation of Multilingual Text Simplification for the Mental Health Domain: Exploring Small Language Models — Multilingual medical text simplification with small LMs is challenging, succeeding only in English for ICD-11 content.
• Linguistic Issues in Computational Linguistics for Non-English Languages — English-centric LLM design creates systematic linguistic weaknesses for non-English languages across multiple levels.
• Cross-lingual Relation Extraction with Large Language Models: Zero-Shot, Few-Shot, and Fine-Tuned Evaluation on Romanian — QLoRA fine-tuning significantly boosts cross-lingual relation extraction performance for low-resource languages, outperforming prompt-only LLMs.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.