Logic-Based Parsing with Neural Networks
August 17, 2022
We see how the unary operations of modal logic can happily coexist with the multiplicative rules of intuitionistic linear logic, and how the two together allow us to pratically model a variety of complex linguistic phenomena at the syntax-semantics interface. We discuss the issues inherent to the natural deduction presentation of the logic, and see how they dissipate if one adopts Girard’s proof nets instead. In the resulting framework, we break proof search apart into a sequence of simple, intuitive steps, and translate them to neural operations. Finally, we play around in real time with the end-to-end model and seek out its strengths and weaknesses.