Abstract
Recent advances in the field of machine learning have shown great promise in solving vari- ous software engineering tasks. However, unlike machine learning techniques used in fields such as NLP (Natural Language Processing) where text-based tokens are used as model inputs, in software engineering (SE) structured representations of source code have proven to be more effective for various SE tasks. Despite the findings, structured representations of source code are still underused. In this paper, we propose to define a benchmark that promotes the usage of structured representa- tions of source code as model inputs, via tasks that are explicitly defined towards that goal.