The Effect of Linguistic Typology on Transfer Learning of Morphology
Date
2020
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Producer
Director
Performer
Choreographer
Costume Designer
Music
Videographer
Lighting Designer
Set Designer
Crew Member
Funder
Rehearsal Director
Concert Coordinator
Advisor
Moderator
Panelist
Alternative Title
Department
Tri-College (Bryn Mawr, Haverford, and Swarthmore Colleges). Department of Linguistics
Haverford College. Department of Computer Science
Haverford College. Department of Computer Science
Type
Thesis
Original Format
Running Time
File Format
Place of Publication
Date Span
Copyright Date
Award
Language
eng
Note
Table of Contents
Terms of Use
Rights Holder
Access Restrictions
Open Access
Terms of Use
Tripod URL
Identifier
Abstract
In the SIGMORPHON 2019 shared task 1, multiple teams attempted for the first time to leverage transfer learning to build more accurate models of natural language morphology with small amounts of target language data, with the intended goal of boosting modeling resources for low-resource languages. It was found that transfer learning could aid the development of computational models for low resource languages and that transfer learning was most effective between genealogically related languages. This study expounds on those findings by testing a much larger number of unrelated language pairs, systematically comparing two model architectures, and examining relationships between model performance and selected linguistic typological similarities of source and target languages. It was found that transfer learning can still afford substantial benefits when source and target language are unrelated, and that transfer learning is most beneficial when source and target language have similar sets of morphologically inflected categories and similar patterns of fusion between those categories, while similarities in inflection shape are not predictive of transfer learning efficacy. This information can be used to select source languages when leveraging transfer learning to improve computational resources for low-resource target languages, especially those without closely related high-resource languages.