A transformer-based parser for Grammatical Framework

Abstract

Grammatical Framework (GF) is a programming language, the purpose of which, among many other things, is to describe natural languages through a common abstract syntax. There is already an algorithm available, to turn natural language into abstract syntax. This algorithm runs in polynomial time, but to produce any output, the input sentence needs to strictly follow grammatical rules. For example, misspelled sentences and sentences containing abbreviations will not be able to be parsed using the existing parsing algorithm. Thus, the objective of this thesis is to develop a transformer-based parser for Grammatical Framework that, given an input sentence, predicts a corresponding abstract syntax that when linearized, resembles the original sentence as closely as possible.