Practical Natural Language Processing for Low-Resource Languages.

Abstract

As the Internet and World Wide Web have continued to gain widespread adoption, the linguistic diversity represented has also been growing. Simultaneously the field of Linguistics is facing a crisis of the opposite sort. Languages are becoming extinct faster than ever before and linguists now estimate that the world could lose more than half of its linguistic diversity by the year 2100. This is a special time for Computational Linguistics; this field has unprecedented access to a great number of low-resource languages, readily available to be studied, but needs to act quickly before political, social, and economic pressures cause these languages to disappear from the Web.

Most work in Computational Linguistics and Natural Language Processing (NLP) focuses on English or other languages that have text corpora of hundreds of millions of words. In this work, we present methods for automatically building NLP tools for low-resource languages with minimal need for human annotation in these languages. We start first with language identification, specifically focusing on word-level language identification, an understudied variant that is necessary for processing Web text and develop highly accurate machine learning methods for this problem. From there we move onto the problems of part-of-speech tagging and dependency parsing. With both of these problems we extend the current state of the art in projected learning to make use of multiple high-resource source languages instead of just a single language. In both tasks, we are able to improve on the best current methods. All of these tools are practically realized in the "Minority Language Server," an online tool that brings these techniques together with low-resource language text on the Web. The Minority Language Server, starting with only a few words in a language can automatically collect text in a language, identify its language and tag its parts of speech. We hope that this system is able to provide a convincing proof of concept for the automatic collection and processing of low-resource language text from the Web, and one that can hopefully be realized before it is too late.