Automatically Inferring Grammar Specifications for Adnominal Possession from Interlinear Glossed Text

Abstract

This thesis presents an update to the AGGREGATION grammar inference project: namely, the ability to automatically infer information about adnominal possession for a given lan- guage. Specifically, I contribute code that begins by examining interlinear glossed text (IGT) instances that have been enriched with information about word alignment, part-of-speech tags, and dependency structures; records information about how those IGT handle adnomi- nal possession; consolidates this information; and finally prepares it to be added to the output grammar specification in accordance with the format expected by the adnominal possession library in the LinGO Grammar Matrix, which in turn can be used to create machine-readable grammars. In this thesis, I describe my contribution to the AGGREGATION project in de- tail, as well as my data-driven development and evaluation process using data from three development languages (Abui, Tsova-Tush, and Nafsan) and three held-out langauges (Mat- sigenka, Wambaya, and Hiaki). My program successfully infers several different approaches to adnominal possession. Specifically, it adds information to the grammar specification that describes the location of possession-marking and the presence of possessor pronoun affixes, as well as adding morphological information about the person information carried by possessive affixes.