Automatically Inferring Grammar Specifications for Valence-changing Verbal Morphology from Interlinear Glossed Text

Abstract

This work builds upon the AGGREGATION project and extends it by adding an inference module targeting valence-changing verbal morphology. This module automatically collects information that answers the questionnaire of the valence-changing verbal morphology library in the LinGO Grammar Matrix. In this study, I show how the module identifies interlinear glossed text (IGT) items with valence-changing morphology and infers information about various types of valence-changing operations from a given IGT corpus. The module was tested on three development (Abui, Hiaki, and Wakhi) and three held-out (Hixkaryana, Old Javanese, and Yaoyos Quechua) languages, all of which are from different language families. The experiments show that the module is able to successfully infer information about multiple valence-changing operations. For all types of operations, the algorithm identifies the transitivity of the verbal stems; for valence-increasing operations, it collects additional information such as the position of the erstwhile subject/added argument on the complements list, the POS of the added argument, and the predicate. In this study, I also perform an extensive error analysis of the experiments, which reveals the limitations in both my module and valence-changing verbal morphology library.