Formal Models of Analogical Proportions

Nicolas Stroppa (1) , François Yvon (2)

1 National Centre for Language Technology
2 LTCI - Laboratoire Traitement et Communication de l'Information

Natural Language Processing (NLP) applications rely, in an increasing number of operational contexts, on machine learning mechanisms which are able to extract, in an entirely automated manner, linguistic regularities from annotated corpora. Among these, analogical learning is characterized by the systematic exploitation, in a symbolic machine learning apparatus, of formal proportionality relationships that exist between training instances. In this paper, we propose a general definition of these proportionality relationships, based on a generic algebraic framework. This definition is specialized to handle a number of representations that are commonly encountered in NLP applications, such as words over a finite alphabet, feature structures, labeled trees, etc. In each of these cases, we provide and discuss algorithms for answering the two main computational challenges posed by proportionality relationships: the validation of a proportion and the computation of the fourth term of a proportion.

Domaines: Linguistique; Informatique et langage [cs.CL]; Traitement du texte et du document

Nicolas Stroppa, François Yvon. Formal Models of Analogical Proportions. 2007. ⟨hal-00145148⟩