Multilingual Variety-aware Language Understanding Technology

There are two main challenges in NLP: a) labeled data is scarce; b) language is ambiguous, even humans might not interpret a linguistic expression the same way. While research exists towards solving the first problem (e.g., via transfer learning or multi-task learning), the second aspect is largely ignored, as computational learning is based on an unrealistic
assumption that annotated samples are unambiguous. I argue that in order to advance the field, a deeper step needs to be taken, and computational learning needs to be linked to human processing, digital traces of which have become available recently (e.g., gaze, keystrokes). This project first studies the deeper connection between human cognitive
traces in language production and reading and computational modeling, by exploiting collections of human annotations with rich behavioral meta-data. Together with recent advances on learning under scarce data, and my pioneering work, this has the potential to
establish a new research direction and firmly embed the human perspective in
computational modeling of language understanding (NLU).