RI: Small: Modeling Lexical Borrowing to Bridge the "Linguistic Divide" in Natural Language Processing

Carnegie Mellon University, Pittsburgh PA

Investigators

Abstract

The rich ecosystem of intelligent, language-aware technologies (e.g., personal assistants, content recommendation, spam detection, etc.) that users of English and other high-resource languages have access to depends on the existence of language-specific data resources. Developing the resources that enable these technologies has usually required a substantial investment -- both monetarily and in terms of trained native speakers -- meaning that without new strategies, most of the 7,000+ languages in the world would likely remain resource-poor and their speakers underserved. This project addresses the problem of bootstrapping linguistic resources required for language technologies in low-resource languages more economically by identifying cross-linguistic correspondences between high- and low-resource languages and projecting resources (e.g., translations, lexical ontologies, and syntactic annotations) accordingly. To identify these correspondences, this work develops computational models of linguistic borrowing, which is the process by which words from a donor language are adapted by speakers of a recipient language as a result of language contact and bilingualism. In addition to enabling the transfer of resources from high- to low-resource languages, being able to identify borrowing enables corpus-based studies of the social factors (power differences between countries, public opinion, and personal attributes such as geographic location, gender, and race/ethnicity) that have been identified as correlates with which words are borrowed. Thus, by observing language change, this work enables changes in social relations to be quantified. Words are not left unchanged by the process of borrowing, and modeling this process is the central challenge to identifying instances of borrowing. Fortunately, the adaptation processes are generally regular and amenable to computational modeling, and this work uses weighted finite-state transducers parameterized with features derived from Optimality Theory (OT). OT-derived features not only provide increased statistical efficiency relative to conventional linguistically naive statistical models but they also provide a new kind of corpus-based verification of some of the central claims of phonological theory. The borrowing model identifies lexical correspondences across dozens of typologically representative language pairs (primary text data is obtained from open resources such as Wikipedia, Twitter, blogs, and online news), enabling projection of resources and development of core natural language processing technologies. Finally, the borrowing model enables instances of borrowed words to be identified in text as it is generated over time, enabling corpus-based sociolinguistic studies.