Semantic Parsing and Generation for the Abstract Meaning Representation

Speaker Name: 
Jeff Flanigan
Speaker Title: 
Postdoctoral Researcher
Speaker Organization: 
UMass Amherst
Start Time: 
Friday, February 22, 2019 - 11:00am
End Time: 
Friday, February 22, 2019 - 12:15pm
Location: 
E2-599
Organizer: 
Marilyn Walker

Abstract:

Uncovering meaning in language and correctly expressing meaning in generated text are key tasks in intelligent language processing.  While deep learning methods have led to breakthroughs in natural language applications, recent research has shown deep learning models tend to rely on surface heuristics and ignore important linguistic and semantic phenomena.  A semantic representation that captures these phenomena for later processing is a possible solution to this problem, and there are ongoing efforts to construct such representations.  In this talk, I give an overview of a recently constructed semantic representation, the Abstract Meaning Representation (AMR), and present parsing and generation algorithms for it.  AMR is a whole-sentence semantic representation which captures relational semantics, or "who is doing what to whom," as a directed graph.  An annotated corpus of over 45,000 sentences paired with their meaning in AMR has been constructed, enabling supervised learning of broad-coverage AMR parsers and generators. After an overview of the semantics represented in AMR, I will present an algorithm for AMR parsing using techniques from combinatorial optimization.  The method uses Lagrangian relaxation combined with an exact algorithm for finding the maximum, spanning, connected subgraph of a graph to produce AMR graphs that satisfy semantic well-formedness constraints.  Next I present an algorithm for generating language from AMR.  The method uses a tree-transducer that operates on a spanning-tree of the input AMR graph to produce natural language sentences, and relies on an argument realization model to overcome data sparsity.

Bio:

Jeff Flanigan earned his Ph.D. from Carnegie Mellon University and is currently a post-doctoral researcher at UMass Amherst.  He works on statistical natural language processing, focusing on semantic parsing and generation.  He built the first AMR parser, which earned Honorable Mention for Best Long Paper at ACL, as well as the first AMR generator.