📄 Kawâkib, a Tool Providing Feedback Between Grammar and Arabic Corpus for the Building of a Theoritical Model (Démonstration de Kawâkib, outil permettant d'assurer le feedback entre grammaire et corpus arabe pour l'élaboration d'un modèle théorique) [in French]

Summary

Kawâkib is an interactive online system for Arabic, shown at the TALN 2014 demo session. It lets a researcher move back and forth between a grammar and a corpus. Its main purpose is to support a "grammar variation" method. The researcher writes or edits a grammar as automata, runs it on texts, studies the output, and revises the grammar. Repeating this loop is meant to lead toward the best algorithm for Arabic morphology, which is central because of Semitic root-and-pattern structure, and for syntax.

The system has two main parts. The first is a library of automata that encode morphosyntactic rules and detect discourse relations. The second is a set of "linguistic radiography" functions that give numeric profiles to arbitrary texts so they can be filtered by relevant criteria.

The paper reports no quantitative evaluation. Instead, the demo walks through grammar–corpus interaction scenarios. The work matters as a practical environment for testing theoretical models of Arabic against real data with little hand-built lexical resource.

Key points

• Morphological analysis of nouns and of triliteral and quadriliteral verbs uses finite-state automata and transducers with minimal linguistic resources. The system displays ambiguities together with their hierarchy.
• Tool words are treated as operator "tokens" grouped into 24 categories. The system provides statistics on these tokens and their categories, and it can locate and store combinations and series of tokens.
• Corpus queries include:
  • roots or root patterns such as R1 X R3;
  • the most frequent roots, with a coverage-threshold calculation;
  • repetitions, temporality markers, reported speech and token-linked discourse relations;
  • regex or automaton patterns that mix words, tokens and logical operators.
• Results from segment parsers can feed tools that synthesize grammar fragments. Users may also modify and instantiate the system's general automaton schemes.
• The system imports Arabic documents in many formats (doc, docx, rtf, pdf) and extracts the text. Authors are from MMSH-CNRS (Aix-en-Provence) and IFAO (Cairo).

Interesting ideas

• I1. Treat grammar as a set of competing variants: run each on the corpus and keep or revise the one whose results fit best, so the model is chosen by evidence rather than fixed in advance.
• I2. Function words act as anchoring operators that drive analysis, so even a minimal-resource analyzer can expose structure through closed-class items and their categories.
• I3. Ambiguity is shown as a ranked hierarchy instead of being flattened into one answer, which keeps the reasoning visible.
• I4. Specific tool words are tied to discourse phenomena such as reported speech, temporality and inter-clause relations.
• I5. Texts get a numeric "radiography" profile built from token and root statistics, used to filter a corpus by discriminating criteria.

For Mova

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

• I1 (learn): Alternative self-induced repair rules or parser settings could be scored side by side on the same corpus, with the winning variant kept and the comparison logged for explainability.
• I2 (tokens): A categorized inventory of English function words could serve as a fallback skeleton when open-class words are unknown or mis-tagged.
• I3 (tree): The parser could keep a ranked list of alternative attachments or tags with reasons, so later stages can revisit a rejected reading.
• I4 (frames): Connectives and reporting markers could directly trigger discourse-relation and speech-act frames, which would make those links traceable to a word.
• I5 (learn): Function-word and construction profiles could screen incoming texts before silver-data agreement and flag out-of-domain material as a negative control.