The Interplay Between Loss Functions and Structural Restrictions in Semantic Dependency Parsing
Robin Kurtz and Marco Kuhlmann. The Interplay Between Loss Functions and Structural Restrictions in Semantic Dependency Parsing. In Proceedings of the Seventh Swedish Language Technology Conference (SLTC), Stockholm, Sweden, 2018.
Semantic dependency parsing, the task of parsing to bilexical directed acyclic graphs representing the semantic structure of a sentence, has recently gained traction due to the ability of these graphs to express language phenomena that cannot be modelled with more restrictive tree structures. The general line of attack on semantic dependency parsing has been to adapt methods and ideas from the more mature field of syntactic parsing, in particular the design of effective inference algorithms for structurally restricted search spaces, and the use of neural networks as the learning component. In this paper we study the interplay between structural restrictions and network loss functions. We uncover a problem that arises when using the usual structured hinge loss in combination with a structural constraint, and propose a modified loss function to address this problem.