CAREER: Modeling Language Evolution via Deep Probabilistic Factorization

University Of California-San Diego, La Jolla CA

Investigators

Abstract

The broad diversity of contemporary and historical languages, dialects, and writing systems presents a daunting challenge for artificial intelligence (AI) systems that must process language data (e.g. systems that automatically recognize handwriting or attempt to translate from one language to another). However, within this great diversity, there are strong patterns of regularity. Historical linguists have shown that many aspects of language evolve over time in accordance with regular patterns of change, including spoken language, spelling, and even the visual appearance of symbols. This project aims to develop novel AI frameworks that can better understand language diversity by automatically analyzing large and diverse datasets consisting of many languages, dialects, and writing systems. The project will result in a collection of new AI systems that track how visual and textual aspects of language evolve over time in order to (1) provide better understanding of how languages change and develop and (2) make downstream AI systems more robust to language diversity. Finally, this research will also support interdisciplinary training of a diverse set of graduate students at the University of California San Diego, as well as the development of interdisciplinary educational modules for high school students interested in AI. This CAREER project will develop a novel computational framework that combines methods from matrix and tensor factorization with deep generative modeling techniques to support analysis of language evolution over a broad range of languages, dialects, and writing systems. The project will create a learning paradigm that (1) incorporates prior phylogenetic knowledge of language history as structured priors, (2) supports efficient approximate inference of historical language forms using neural decoders, (3) is easily portable to a variety of linguistic domains and levels of language representation, and (4) directly analyzes primary data (e.g. images of signs) rather than manually-curated feature lists. Further, the framework will generalize across both visual and textual modalities, allowing for study of the multi-modal nature language evolution -- e.g. scripts evolve through visual change, cognates through phonetic or orthographic change -- and potentially laying the groundwork for future work investigating how script and dialect co-evolve or cultural evolution studies of spoken audio. Finally, the outcomes of each of several applied studies may lead to new evidence for specific historical and paleographic hypotheses. 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.