Séminaire de Recherche en Linguistique

This talk presents two different models with regard to agreement with coordination structures, where more than one agreement strategy exists: resolution rules ("the lawyer or the assistant are coming to the meeting"), closest conjunct agreement ("the lawyer or the assistant is coming to the meeting"). In the first model, we test the linguistic generalizations learned by neural networks. Even though neural networks have achieved very high accuracy in various types of agreement, such as long distance subject-verb agreement (cf. Gulordava et al. 2018), but they use "a bag of features" strategy for agreement with coordination phrases, which is far from achieving human performances. 

     The second model is inspired from harmonic grammar in phonology and phonetics (cf. Hayes and Wilson 2008). The wellformedness of a structure is correlated to the sum weighted constraint violations. The model is trained from a limited set of acceptability data, and can easily generalize to new agreement patterns. Furthermore, we demonstrate that such a model not only unifies various aspects of agreement (nominal/verbal domain, number and gender feature, and and or coordination), but also can capture some typological generalizations (cf. agreement hierarchy by Corbett 1991).