Learning to Reason for NLP

Despite the recent advances in natural language processing and human-level performances of state-of- the-art neural models on common benchmarks, recent models lack various reasoning capabilities. For in- stance, they struggle on datasets that require performing coreference or arithmetic reasoning [Wu et. al, 2021, Moosavi et al., 2021]. In this regard, we explore two different directions: (1) developing innovative models that have improved reasoning capabilities to solve the existing reasoning-aware challenge da- tasets, and (2) creating new benchmarks for less-explored reasoning skills.