Don't Gloss Over the Gloss: Using Grammatical Features to Improve Morphotactic Inference in the AGGREGATION System

Abstract

In the field of linguistics, language data is often presented in the form of interlinear glossed text (IGT). Due to the large amount of work and language analysis required to produce IGT, as well as its widespread use and somewhat standardized format, IGT data is a rich source of structured linguistic information. The AGGREGATION project is an ongoing effort to employ this information for the automatic inference of implemented grammars. One of the perennial challenges for automatic grammar inference is producing grammars that not only make correct predictions about language data, but also whose structures are intelligible to linguists. In this regard, grammars produced by the AGGREGATION system often fall short, particularly when it comes to languages' morphotactic systems, which are modeled as a series of position classes. This thesis works towards remedying this issue by exploring the ways in which the grammatical feature information in IGT gloss lines can be better utilized for morphotactic inference. By reorienting the inference process towards creating position classes based on grammatical features, the quality of the resulting grammars was improved without impacting their performance. These changes were evaluated using datasets representing six diverse languages, each from a different language family and geographical location. Three of these datasets (comprising of data from Wambaya, Lezgi, and Manipuri) were consulted during the development process, while the other three (comprising of data from Wakhi, South Efate, and Hiaki) were held out and used in the final evaluation.