📄 SIMPLE: A General Framework for the Development of Multilingual Lexicons

Summary

SIMPLE (LE-SIMPLE), an EC-funded Language Engineering project, sets out to build wide-coverage semantic lexicons for 12 European languages under one shared model so they stay consistent and can later be linked across languages. It extends the LE-PAROLE morphological and syntactic lexicons with a semantic layer for about 10,000 word meanings per language, partly drawing on the PAROLE corpora. The model draws on the EAGLES recommendations and an extended Generative Lexicon. Word senses are encoded as Semantic Units (SemUs). Each SemU receives a type from a common ontology organised by the four Qualia roles, and its encoding is guided by language-independent templates. The model covers argument structure, selectional restrictions, event type, linking to PAROLE subcategorization frames, regular polysemy and cross-part-of-speech relations. The paper reports no experimental figures. Its main contribution is the framework itself and a harmonised, SGML/GENELEX-conformant resource that ELRA will distribute. The work matters because it tests lexical-semantic theory at multilingual scale and gives NLP applications a semantic resource that complements EuroWordNet.

Key points

• Languages covered: English, French, German, Spanish, Italian, Portuguese and Catalan, together with Greek, Dutch, Danish, Swedish and Finnish. Each lexicon holds roughly 10,000 senses built on top of LE-PAROLE.
• About 800 EuroWordNet "base concepts" were used as the shared core sense set. Through the EuroWordNet Interlingual Index, these give automatic cross-language links among all 12 languages.
• The ontology has two layers:
  • a Core Ontology of top, shared types;
  • a Recommended Ontology of finer-grained types.
• The top levels follow the Qualia roles (Telic, Agentive, Constitutive) alongside Entity.
• The model distinguishes features, which take closed sets of values, from relations between SemUs, which are organised hierarchically. Example relations are Is_a_member_of (senator to senate), Is_a_part_of (head to body) and Used_for (eye to see).
• All data is encoded in SGML following the GENELEX DTD. External experts are to validate the lexicons by sampling, and the type system is to be checked through its use in selectional restrictions.

Interesting ideas

• Templates as blueprints. Each type is tied to a template that states which information defines the type and which is optional. This makes the ontology both a formal specification and a practical encoding guide, which helps keep many annotators or teams consistent in any multilingual resource.
• Defining abstract nouns by parts rather than parents. Senses like "part" or "group" sit under a Constitutive node and are characterised by meronymic links, since their hypernyms say little. This could help model relational and abstract nouns that resist ordinary taxonomies.
• Regular polysemy versus homonymy. Senses in a systematic alternation are linked by a named relation, such as BuildingInstitution for the two senses of school. Unrelated meanings, as in bank, are simply left unlinked. Annotating polysemy this way lets you mine and compare its patterns across languages.
• A three-layer syntax–semantics interface. Syntactic frames, predicate-argument structure and a correspondence layer are kept apart. As a result, destroy and destruction can share one predicate and its selectional restrictions while linking to syntax differently. The design suits derivational families in morphologically rich languages such as Ukrainian.
• Anchoring to a shared core sense set. Building every language around the same base concepts gives cross-lingual alignment almost for free, which is a cheap strategy when starting new parallel lexicons.