📄 NEMLAR - An Arabic Language Resources Project

Summary

NEMLAR (Network for Euro-Mediterranean LAnguage Resources) is an EU-funded project running 2003–2005. It targets a basic problem for Arabic language technology: there are too few usable, standardised and shareable language resources (LRs) and tools. The project brings together European centres and partners in Jordan, Morocco, Egypt, Lebanon, Tunisia, and the West Bank and Gaza Strip. It has four strands of work:

• mapping the players in the region and the resources they hold, recorded in a knowledge base;
• asking industry which resources it lacks;
• defining a BLARK (Basic Language Resource Kit) for Arabic, the first such definition for the language;
• planning pilot projects to update or create resources.

The survey found that lexica with rich morphology and morpho-syntax exist, along with POS taggers and text-to-speech products. Valency and semantic lexica are scarce, and no freely available large grammar or parser exists. Many resources held by companies stay closed. The paper also reports how resources are validated and how willing their owners are to share them. It sets out an ELRA-style validation approach as the route to raising resources to a reusable standard. Dissemination runs through a website, a newsletter and a 2004 conference.

Key points

• The knowledge base lists 35 institutional players (22 in Arabic-speaking countries: Egypt 8, Lebanon 4, Jordan 3, Palestine 3) and about 20 individuals. Key companies include Sakhr, IBM Egypt and RDI, with Systran and France Télécom active in Europe.
• At the time of the survey, LDC distributed 15 Arabic LRs and ELRA distributed 3. No semantic lexicon comparable to Euro-Wordnet existed.
• Validation practice among respondents:
  • 67% validate internally only;
  • 19% use external validators;
  • 13% do not validate at all.
• Distribution attitudes:
  • over half of respondents would share their resources under a standardised negotiated agreement;
  • 19% refuse, for legal (9%), commercial (6%) or strategic (4%) reasons.
• The Arabic BLARK differs from the Dutch original in three ways:
  • it includes a diacritizer;
  • it describes availability in finer detail than Binnenpoorte 2002;
  • it adds quality, including adherence to standards, as a separate dimension.

Interesting ideas

• I1. Validation means checking a resource against its own documentation; the formal check asks whether exactly the documented attributes and values appear, no more and no fewer.
• I2. Formal validation (technical and schema checks) is kept separate from content validation, which needs a speaker with strong command of the language to judge whether entries are coded correctly.
• I3. A minimal resource kit is defined per language, so a component that one language can skip may be essential for another (a diacritizer for Arabic but not for Dutch).
• I4. Each kit component is rated for both availability and quality, because a resource that exists but ignores standards does not really close the gap.
• I5. Inventorying lexica by annotation layer exposed a clear gap: morphology was well covered, while valency and semantic information were rare.

For Mova

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

• I1 (tree): Each treebank or converted UD dialect could carry a declared label and feature inventory, with an automatic check that flags undeclared or never-used deprels and features.
• I2 (learn): Silver data could pass a cheap automatic schema gate first, then a separate content audit by LLM-teacher tiers or human spot checks, with each gate's error rates reported separately.
• I3 (world): Mova's compiled knowledge layers could be described as a minimal kit (verb classes, measures, ontology, knowledge graph), listing which parts each target domain actually requires.
• I4 (learn): Mova's component dashboard could score every resource and model on two axes, presence and verified quality, so that weakly checked parts are not counted as finished.
• I5 (frames): A per-layer coverage audit of the lexicon (morphology, then valency, then VerbNet roles) would show where frame building lacks support and where effort should go first.