Abstract

One of the central problems in language technology is the realization of an accurate mapping between natural language utterances and in-depth meaning representations. In this talk, I will discuss methods for this bi-directional mapping which exploit graph-centric representations, formalisms, algorithms and neural networks. I will first introduce a neural graph rewriting framework to model syntactico-semantic composition which combines the strengths of Hyperedge Replacement Grammar (HRG), for knowledge-intensive learning, and Graph Neural Networks, for data-intensive learning. I will then discuss two fundamental problems in this framework: semantic graph parsing and parsing semantic graphs, i.e. computing all possible / the best derivation(s) of a given string or graph. In particular, I will demonstrate that exact graph parsing can be efficient for large graphs and with large grammars. With the ability to enumerate every derivation of a surface string or a semantic graph, we are ready to build practical language understanding and generation systems. I will report on neural systems that achieve state-of-the-art accuracy for English Resource Semantics. I will conclude by laying out a few ideas for future work on meaning representation—mediated Natural Language Processing.

Link to join: https://cl-cam-ac-uk.zoom.us/j/91253900399?pwd=SU5TNnpYdDlQbzQ4SEVPVWVWa0Nldz09

A recording of this talk is available to members of the Department at the following link: https://www.cl.cam.ac.uk/seminars/wednesday/video/