What Do Neural Networks Actually Learn, When They Learn to Identify Idioms?

To date, Neural Networks (NNs) have been employed to carry out the automatic classification of various kinds of figurative expressions, like idioms (Bizzoni et al.,2017b) and metaphors (Do Dinh and Gurevych,2016; Bizzoni et al.,2017a;Rei et al.,2017). It is common knowledge that metaphors (e.g., my job is a jail) reflect a transparent mapping from concrete examples ina source domain (e.g., the physical confinementof a jail) to abstract concepts in a target domain (e.g., the psychological constraints and tediousness of a job) (Lakoff and Johnson,2008), while idioms (e.g., buy the farm ‘to pass away’, shootthe breeze ‘to chat idly’) synchronically appear as a rather heterogeneous class of semantically non-compositional multiword units that all in all exhibit greater lexicosyntactic rigidity, proverbiality and emotional valence with respect to literal expressions (Nunberg et al.,1994;Cacciari,2014)