📄 Creation and Evaluation of Extensible Language Resources for Maltese

Summary

Maltese lacked large computerised language resources until the University of Malta launched its Maltilex Project. Maltese is a mixed language with an Arabic-based morphology and syntax, a large Romance vocabulary and loosely integrated English borrowings. This paper describes how the project built an extensible resource infrastructure for it.

A weakly supervised method called the Lexicon Structuring Technique (LST) automatically grouped word forms from a raw corpus under lemmas. This produced the core of a computational lexicon. The lexicon sits in a relational database whose fields map onto XCES, and new functionality can be added through an extension API.

For interoperability, the author proposes a web-services model. Each resource is wrapped by a SOAP server and described with WSDL and UDDI, and a conversion layer hides the resource's internal format.

The paper also introduces two evaluation measures, L-Measure and C-Measure. They suit minority languages because no gold-standard resource is needed for comparison. The work matters as a template for building and exposing resources for under-resourced languages from scratch.

Key points

• The starting corpus was roughly 3,000 Maltese texts with over 2 million word forms. LST produced a lexicon of about 80,000 lemmas, and the paper also reports around 60,000 unique word forms.
• LST draws on bioinformatics and data-mining principles. It assigns each lemma a headword chosen from one of its grouped word forms.
• The core lexicon stores orthographic and phonological information and maps almost directly to XCES. The project also built an XCES-compliant tagset for Maltese (Gatt, 2001).
• Resources are exposed through a SOAP server described in WSDL, with UDDI for registration. An example service, LRCoreService, offers the function getEtymology, which returns language codes for a word's etymological source.
• The C-Measure ranges from 1 (full coverage) down to −γ (no coverage). γ reflects confidence in corpus quality and would be set to 0.95 in the paper's example. The L-Measure combines recall and precision with a harmonic-mean formula.

Interesting ideas

• I1. Use the lemma rather than the individual word form as the unit of reference: store irregular forms explicitly and generate the regular ones by rule, which keeps the lexicon compact and lets relations attach once per lemma.
• I2. Keep data in a fast relational store and convert it to XML only on export, filtering with SQL before conversion instead of running costly transformations over everything.
• I3. Wrap each resource behind a two-way conversion layer that translates its native format into a standard core interface and back, so other tools can both read and update it.
• I4. Measure lexicon coverage against several reference dictionaries without a gold standard, and weight each mismatch by how far those references can be trusted.
• I5. Assign a service to each major component and expose its smaller operations as functions, so the number of top-level services stays manageable.

For Mova

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

• I1 (tokens): The en lexicon could keep only exceptional inflections per lemma and derive the rest, so lemma-level facts don't need duplicating across forms.
• I2 (world): Compiled knowledge from WordNet, FrameNet and Wikidata could live in an indexed store, with interchange formats produced only for the requested slice.
• I3 (tree): UD dialect conversion could be built as reversible adapters around each treebank, making round-trip loss directly testable.
• I4 (learn): Silver-data evaluation could split disagreements into "only ours" and "only the witness" buckets, with weights tied to witness reliability.
• I5 (answer): Readers and students could expose a few coarse entry points with fine-grained calls inside, which keeps explanations and APIs easy to navigate.