Discovering and Demonstrating Linguistic Features for Language Documentation

点击翻译按钮获取中文摘要

Documenting endangered languages is matter of great urgency, but is also a time-consuming process. Annotation and curation of speech and text data, searching for interesting, prototypical, or atypical entries, and marshalling examples for pedagogy or publication are still mostly manual processes. This project aims to speed these processes by (1) creating better tools for the automated analysis ofsmaller and partially-annotated corpora, which have potential to reduce the amount of time required by linguists to manually annotate these corpora, and (2) creating better methods for linguists to browse their collected data and answer questions about the characteristics of the language at hand. The intellectual contribution of this proposal will lie in the development of new computational methods for natural language processing (NLP) for endangered languages, and their evaluation, both in controlled environments and as a tool for linguists in the field. It will also have broader impact in the creation of new tools and standards for linguistic documentation,increased collaboration between linguists and computer scientists, and training of a graduate student in the technologies and practices necessary to move this collaboration forward. The training component will increase the STEM workforce capacity in computational linguistics, important given the need for more advanced tools in working on languages that are underdocumented and spoken in countries that are key to national interests. As a specific methodology to realize this vision, this project focuses on recent development of massively multilingual NLP models based on neural networks. These methods work by creating NLP using data from a large number of languages, then using the information gleaned from these languages to improve the accuracy of processing on a new language with a paucity of training data. Within this framework, three major research questions will be examined: (1) How can these techniques be efficiently applied to very-low-resource languages,especially those in the early stages of text collection? (2) What methods can be used to move beyond sentence-by-sentence analyses, and synthesize information about the entirety of the language to propose a simple grammatical specification? (3) Is it possible to provide examples that support typological predictions for a linguist to read and learn more about the nuances of the language they are analyzing? All three of these research questions will be examined in a rigorous process of devising methods, testing on existing data sets for well-resourced languages, and finally deployment to field linguists to examine how they improve the efficiency or accuracy of the language documentation process.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.