MURAD: A Large-Scale Multi-Domain Unified Reverse Arabic Dictionary Dataset

AI summary

Overview

• Research area: Arabic natural language processing (NLP) and computational lexicography, specifically the construction of a reverse dictionary resource.
• Technical level: Intermediate. The methods are described conceptually (text extraction, OCR, normalization, descriptive statistics) and the paper is a dataset/resource paper rather than an algorithmic one, so it is readable without deep modeling background.
• Scope: The paper introduces MURAD, an open Arabic dataset of 96,243 word–definition pairs drawn from 17 curated reference works across 13 domains, intended for reverse dictionary modeling, semantic retrieval, and related Arabic lexical-semantics research.

What This Paper Is About

Arabic has a vast vocabulary spanning scientific, religious, and literary domains, but large-scale lexical resources that link Arabic words to precise formal definitions are limited. Reverse dictionaries (systems that let users search by meaning rather than by word) are well established for languages such as English, French, and Chinese, but comparable Arabic resources have lagged behind because of Arabic's morphological richness, diglossia, and orthographic ambiguity. The goal of this work is to build and release MURAD (Multi-domain Unified Reverse Arabic Dictionary), a large, multi-domain, standardized Arabic word–definition dataset that can support both human reading and computational modeling.

Key Contributions

1. The largest curated Arabic word–definition dataset described in the paper: 96,243 word–definition pairs, which the authors state is developed in accordance with established lexicographic standards and validated through expert review.
2. Comprehensive multi-domain coverage: the dataset spans classical Arabic, linguistics, Islamic studies, and scientific and technical terminology, distributed over 13 domains mapped into 4 high-level categories and drawn from 17 curated sources.
3. A hybrid, fully reproducible extraction pipeline: direct text parsing, optical character recognition, and automated reconstruction, with the complete processing workflow released as a public code library.
4. An open and publicly released resource: the dataset is available on Hugging Face under CC BY 4.0, with documentation, to support reproducible research on Arabic lexical semantics.

Main Findings

• Dataset size and structure: MURAD contains 96,243 curated (Word, Definition, Reference) triplets, stored as a single plain-text, comma-delimited, UTF-8 encoded CSV file along with a README describing the dataset and its sources.
• Corpus statistics: the dataset contains 1,482,322 total word tokens across all text (words plus definitions), a corpus-wide unique vocabulary of 95,746 word types, and an average definition length of 13.8 words.
• Definition length distribution: most definitions contain between 10 and 20 words, with a median of 13 words and a mean of 13.8 words; the distribution is right-skewed, with a small number of extended definitions reaching up to 100 words. The authors describe this as authentic variation in definitional style rather than noise.
• Domain balance: General terms make up the largest share at 38.41%, followed by Scientific at 29.15%, Islamic at 18.57%, and Linguistic at 13.87% of total content.
• Comparison with existing resources: MURAD is described as larger and broader than prior Arabic resources. KSAA-RD is reported as containing 58,000 words, without formal lexicographic guidelines and with limited domain breadth; KSAA-CAD focuses on concept–attribute descriptions rather than dictionary-style definitions and is smaller and narrower; Azhary provides structured semantic relations such as synonymy and hypernymy but lacks definitional text. In Table 1, KSAA-CAD is listed with 58,010 Arabic and 4,355 English entries, and Azhary with 26,195 words / 13,328 synsets, while MURAD has 96,243 curated entries from 17 curated sources with a structured lexicographic standard and fully open access.
• Multilingual positioning: the SemEval-2022 Task 1 CODWOE shared task explored bidirectional mapping between definitions and word embeddings across English, French, Spanish, and Chinese; MURAD complements it by extending definition–embedding alignment to Arabic, which CODWOE did not include.
• Source contribution counts: the 17 sources contributed the following numbers of definitions — Al-Kafawi's Dictionary of Universals 14,476; Al-Jurjani's Book of Definitions 1,399; Dictionary of Chemistry Terms 4,468; Dictionary of Machine Learning Terms 1,758; Dictionary of Mathematical Terms 7,808; Dictionary of Physics Terms 5,081; Dictionary of Arabic Measurement Terms 550; Dictionary of Psychology Terms 4,157; Dictionary of Mechanical Engineering Terms 1,569; Book of Terminology in Arabic Sciences 13,181; Encyclopedic Dictionary of Applied Linguistics Terms 13,350; Dictionary of Islamic Jurisprudence Terms 9,964; Dictionary of Electrical, Electronic, and Communication Engineering Terms 1,422; Dictionary of Scholars' Terminology 7,907; Encyclopedia of Faith Terminology 4,160; General Terminology Dictionary 3,750; SDAIA Data and Artificial Intelligence Glossary 1,243. These sum to the reported total of 96,243.
• Not reported: the paper text provided does not include the specific example words and definitions listed in Table 5, so no sample entries can be quoted here. No model training results or reverse-dictionary accuracy benchmarks are reported; the validation is structural and descriptive.

Methodology in Plain English

The authors collected publicly available Arabic reference works and specialized glossaries covering Islamic studies, linguistics, mathematics, chemistry, physics, psychology, and engineering. Most sources were already digital PDF or structured HTML, but some existed only as printed volumes.

For the printed material, high-resolution scans were processed with Mistral OCR to produce machine-readable text, and GPT-4o was used for semantic text extraction and structuring to pull out clear, consistently formatted word–definition pairs. For digital PDFs and HTML, automated scripts extracted Arabic text line by line while preserving character order and diacritics, and regular-expression filters kept valid Arabic Unicode ranges while stripping out pagination, tables, and headers.

After extraction, lines were segmented and joined into continuous definition entries so that each term was paired correctly with its definition. Preprocessing also normalized Arabic character variants and unified punctuation. Duplicate and incomplete records were detected automatically and removed to keep the resource internally consistent. The cleaned pairs were then merged, deduplicated, and exported as a UTF-8 CSV indexed by three fields: word, definition, and ref (the source identifier).

Validation covered three areas: data integrity (automated checks that every row had a valid non-null Arabic term with its definition and valid UTF-8 text), linguistic consistency (word counts and descriptive statistics on definition lengths), and domain balance (distribution of definitions across the four major source categories). All validation and analysis used open-source tools — pandas, numpy, and matplotlib — with scripts published in the dataset's GitHub repository, and all reported figures and statistics were generated directly from the released dataset without additional filtering.

Why This Matters

Impact on research. The paper positions MURAD as filling a gap in Arabic NLP: large-scale, semantically aligned datasets connecting words to formal definitions, which the authors argue are essential for meaning-based tasks such as semantic retrieval and word-sense disambiguation. It also extends definition–embedding alignment work (previously done for English, French, Spanish, and Chinese via CODWOE) into Arabic, and complements ontology-style resources such as Azhary that lack definitional text.

Real-world applications (as described by the authors):

• Reverse dictionary systems that mitigate the "tip of the tongue" phenomenon, letting users retrieve a word from its meaning.
• Term selection and terminological consistency in high-stakes domains such as legal Arabic, where the same concept may be expressed with different terms across countries or schools of jurisprudence, and in engineering, where borrowed English or French concepts produce competing Arabic translations.
• Educational tools, including definition modeling and dictionary-style gloss generation or interpretation.
• Semantic retrieval, word-sense disambiguation, embedding evaluation, interpretable model analysis, and cross-lingual semantic alignment studies.

Industry relevance. Practitioners working on Arabic language technologies — search and retrieval, terminology management in law and engineering, educational content, and Arabic NLP model development — gain a freely available, standardized corpus they can plug into existing pipelines via CSV. The public code library also lowers the barrier for organizations wanting to build similar term–definition datasets from their own PDF collections.

Future Directions

• Benchmarking reverse-dictionary models on MURAD: no baseline reverse-dictionary or definition-modeling results are reported, leaving head-to-head comparisons against Arabic resources such as KSAA-RD as an open task.
• Dataset expansion and auditing: the authors explicitly frame the open validation framework as facilitating future benchmarking, auditing, and dataset expansion by the research community.
• Addressing the morphological and orthographic challenges the paper identifies: consistent lexical annotation for a morphologically rich, diglossic language remains an open problem the resource only partially addresses.
• Extending coverage beyond the current 13 domains and 17 sources: the domain distribution is uneven (General 38.41% versus Linguistic 13.87%), and additional reference works could rebalance and broaden lexical coverage.

Target Audience

Researchers and practitioners in Arabic NLP and computational lexicography, especially those working on reverse dictionaries, semantic retrieval, definition modeling, word-sense disambiguation, and embedding evaluation. It is also relevant to terminologists and translators in legal, engineering, and scientific Arabic contexts, and to educators or developers who need a structured, openly licensed Arabic word–definition resource for building tools.

Authors’ abstract

Arabic is a linguistically and culturally rich language with a vast vocabulary that spans scientific, religious, and literary domains. Yet, large-scale lexical datasets linking Arabic words to precise definitions remain limited. We present MURAD (Multi-domain Unified Reverse Arabic Dictionary), an open lexical dataset with 96,243 word-definition pairs. The data come from trusted reference works and educational sources. Extraction used a hybrid pipeline integrating direct text parsing, optical character recognition, and automated reconstruction. This ensures accuracy and clarity. Each record aligns a target word with its standardized Arabic definition and metadata that identifies the source domain. The dataset covers terms from linguistics, Islamic studies, mathematics, physics, psychology, and engineering. It supports computational linguistics and lexicographic research. Applications include reverse dictionary modeling, semantic retrieval, and educational tools. By releasing this resource, we aim to advance Arabic natural language processing and promote reproducible research on Arabic lexical semantics.