Language variation: challenges and future avenues
Recent developments in deep learning have shown striking performance improvements on a wide array of natural language processing (NLP) tasks. Despite the great progress, current approaches to NLP typically assume language is homogeneous, and thus fail to account for the intra- and extra-linguistic variations encoded in language, such as domains, genres, languages, and social factors. The main consequence is a dramatic drop in performance on out-of-distribution data, and unpredictable behaviors on new distributions, including the magnification of harmful stereotypes, and unfair and discriminatory decision making. In this talk, I will dig into the topic, introducing the theoretical notion of the variety space and the challenges which language variation entails. Current efforts and future avenues for research will be discussed, including transfer learning approaches, and the need for living benchmarks and transparent model and data statements.
Alan Ramponi is a PhD candidate in natural language processing at the University of Trento, Italy. He was a fellow at the Microsoft Research COSBI centre, Italy (2017-20), and a visiting PhD fellow in the NLPnorth group at the IT University of Copenhagen, Denmark (2019-20). He received the MSc in computer science cum laude from the University of Trento in 2017. His research focuses on language variation, and specifically in making natural language processing (NLP) robust and ultimately aware of variation across domains, social factors, and languages.
When: 09/04/2021 at 11.00
Where: https://unito.webex.com/webappng/sites/unito/meeting/info/910eaf7ad0534d1ba92c5dde0a66a9a7