Compositional semantic parsing across graphbanks
Alexander Koller, Saarland University

Over the past few years, quite a few semantically annotated corpora have become available. While many of these use graphs as their semantic representations, there are considerable differences in the design of these graph annotations. In my talk, I will first look at a number of semantic graphbanks and discuss some of these differences.

Because of these design differences, most semantic parsers that are accurate for one graphbank do not perform well on another. For instance, a semantic dependency parser which assumes a one-to-one correspondence between words and nodes will struggle on AMR. I will show how to build a semantic parser which works well across a number of substantially different graphbanks. Our semantic parser predicts the compositional structure of a semantic representation using a neural dependency tree parser and then evaluates it to a graph by evaluating the dependency tree in a graph algebra. This allows us to generalize across graphbanks.

I will conclude by discussing phenomena which cause mismatches between the compositional structures of different graphbanks, which may help us understand the differences in design decisions even more clearly.

Bio:

Alexander Koller is a Professor of Computational Linguistics in the Department of Language Science and Technology at Saarland University. He received his PhD from Saarland University for research on graph-based underspecification methods for scope ambiguities, and has also worked at Columbia University, the University of Edinburgh, and the University of Potsdam. His research interested span a broad range of topics in computational linguistics, including semantics, parsing, and dialogue.