The Relationship between Parsing and Verbal MWE Identification: It's Complicated

It is commonly assumed that verbal multiword expression (VMWE) identification and syntacticparsing can mutually benefit from one another, motivating a wide range of joint and syntax-awareapproaches. In this paper, we systematically re-examine this assumption and show that the relation-ship between parsing and VMWE identification is far from straightforward. Using gold-standarddependency and VMWE annotations from French PARSEME 1.2 and English STREUSLE, we evalu-ate multiple ways of incorporating syntactic information into neural VMWE identification models.These include multi-task learning with shared representations, uncertainty-based dynamic loss weight-ing, supervised contrastive learning over dependency relations, and explicit dependency-relationembeddings concatenated to contextualized token representations. While dynamic loss weightingimproves over naive multi-task training, none of the joint or syntax-enhanced approaches consistentlyoutperform a strong MWE-only baseline. In several settings, syntactic supervision substantiallydegrades performance and leads to unstable training behavior. Our results suggest that, despite strongtheoretical motivation, current neural models struggle to exploit explicit syntactic information forVMWE identification, and that simpler task-specific architectures remain more effective.