Learning to Unlearn in Lattices of Concepts: A Case Study in Fluid Construction Grammars

This paper outlines a couple of lattice-based (un)learning strategies proposed in a recent development of unification-based grammars, namely the Fluid Construction Grammar (FCG) setup. These (un)learning strategies are inspired by two linguistic phenomena occurring in a dialect spoken in the Banat area of Romania. Children from that region -- where influences produced over centuries by Serbian, a Slavic language, are obvious -- learn in school the modern Romanian language, which is a Romance language. This particular setup offers us the possibility to model in FCG a two-step learning process: the first step is that of learning a (perfective) verbal aspect similar to the one already presented by Kateryna Gerasymova in her MSc thesis, while the second one is concerned with un-learning (or, learning another linguistic "construction'' over) this verbal aspect. Thus, the interesting issue here is how learning could continue beyond learning the verbal aspects. We will first give linguistic facts, after which we will outline the way in which FCG could model such a linguistic process. From the computational point of view, we show that the heuristics used in this grammar repairing process can be automatically derived since the meanings associated to words and phrases are organized in a lattice of feature structures, according to the underlying constraint logics. We will later discuss the case of another verbal marker in the dialect spoken in Banat. It will lead us to sketch a composite, quite elaborated (un)learning strategy.