📄 An Online Readability Leveled Arabic Thesaurus

Summary

Arabic is hard for learners. It has rich morphology, and words are usually written without short vowels, so a single written form can map to several unrelated lemmas. Few dictionaries in any language mark how difficult a word is, and none in Arabic did so before this work. The authors built a web thesaurus that combines three existing resources. Camel Tools with its CALIMA Star analyzer turns an input into candidate lemmas. Arabic WordNet (AWN) supplies lexical relations. The Arabic Readability Lexicon assigns each lemma one of five difficulty levels. Users can search by inflected word, lemma, root or English gloss. For each candidate lemma, the system lists its gloss, root and level, plus synonyms, antonyms, hypernyms, hyponyms, related expressions and phrases containing the lemma. Every displayed item links to a new search. The main engineering work was aligning AWN's lemma and POS conventions with those of Camel Tools. The tool gives teachers and learners access to Arabic NLP research, and user feedback is expected to reveal coverage gaps and errors.

Key points

• The demo is a Flask (Python) back end with a light JavaScript front end. It uses 11 look-up tables, nearly all keyed by a combined lemma#pos string.
• CALIMA Star extends SAMA and covers more than 40 thousand lemmas, each tied to a POS, a root and an English gloss.
• AWN contains 16,066 entries, of which 5,036 are multiword phrases, grouped into 14,284 synsets. AWN relations outside the four standard types are merged into a single "related" category.
• The Arabic Readability Lexicon has 26,578 lemmas that language professionals rated on five levels, running from Grade 1 (age 6) to specialist (age 15+).
• AWN lemmas were matched to Camel Tools lemmas offline. Entries were run through MADAMIRA, best candidates were picked by Levenshtein distance, and hand-written rules mapped POS categories between the two resources.

Interesting ideas

• I1. Using one combined lemma-plus-POS key across every table keeps the resources joinable and separates homographs that differ only in part of speech.
• I2. Two lexicons with different lemma conventions can be reconciled by running one lexicon's entries through the other's analyzer, then ranking candidates by edit distance and constraining them with explicit POS mapping rules.
• I3. Instead of committing to one reading, the interface shows every morphological analysis of an ambiguous input and lets relations unfold separately under each one.
• I4. A difficulty level is attached to every lemma, including the related words that are suggested, which makes it possible to pick a simpler synonym on purpose.
• I5. Synonyms are the union of co-members across all synsets of a lemma, and antonymy is defined between synsets and then projected down to their lemmas.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (world): Key compiled WordNet/VerbNet/Wikidata entries on lemma#UPOS so that noun and verb senses of the same string never collide in the knowledge graph.
• I2 (world): Apply this recipe to align FrameNet and WordNet lemma spellings with crate en lemmas, and log each mapping rule so the alignment stays explainable.
• I3 (tokens): Pass all candidate lemma analyses downstream with their sources, so that later stages, rather than the tagger alone, settle the ambiguity.
• I4 (back): During generation or simplification, prefer the lowest-level synonym from the frame's lexical options when writing for young readers or school-course tasks.
• I5 (world): Use the same projection to derive lemma-level antonym and synonym facts from WordNet synset relations, keeping a trace of which synset licensed each fact.