Towards Cross-lingual Representation of Prototypical Lexical Knowledge

Résumé 0

In order to be concretely effective, many Natural Language Processing (NLP) applications require the availability of lexical resources providing varied, broadly shared, and language-unbounded (i.e., not limited to any specific language or linguistic system) lexical information. However, state-of-the-art knowledge models typically focus on specific levels of semantic analysis rather than adopting such a comprehensive and cross-lingual approach to lexical knowledge. This is often due to the theoretical paradigms on which such resources are based, each addressing the semantic phenomenon from a (de)finite perspective. In this paper, we first suggest a maximalist approach to lexical semantics to pursue through the idea of semantic prototype and linguistic representativeness as easily applicable to textual corpora. Starting from this conceptual framework, we thus propose a novel corpus-based automatable methodology for knowledge modeling based on a multilingual word alignment mechanism. This model enables the retrieval and encoding of prototypical, language-unbounded, and naturally disambiguated lexical knowledge in the form of diversified conceptual links between words and their senses. Results from a simple implementation of the proposal show relevant outcomes that are not found in other resources. Finally, different application opportunities of the proposed model will be presented.