Parsing And Generation For The Abstract Meaning Representation
Uncovering meaning in language and correctly expressing meaning in generated text are key tasks in intelligent language processing. In this talk, I introduce parsing and generation algorithms for the Abstract Meaning Representation (AMR). AMR is a whole-sentence semantic representation which captures relational semantics, or "who is doing what to whom," as a directed graph. Recently, an annotated corpus of over 45,000 sentences has been constructed, enabling training of broad-coverage AMR parsers and generators.
First, I will present a method for parsing into AMR using techniques from combinatorial optimization. The method consists of two stages, which first identify concepts and then add relations between them using a maximum connected subgraph algorithm, utilizing Lagrangian relaxation to enforce constraints on well-formedness.
Next I present a method for generating language from AMR using weighted tree transducers. The method first converts the AMR graph to a tree and which is then transduced to a string, and relies on discriminative learning and an argument realization model to overcome data sparsity.
Jeff Flanigan is a Ph.D. candidate at Carnegie Mellon University in the Language Technologies Institute. He works on statistical natural language processing, focusing on semantic parsing, generation, and machine translation. He built the first AMR parser, which earned Honorable Mention for Best Long Paper at ACL 2014, as well as the first AMR generator. His undergraduate degree is from the University of California, Santa Barbara, and he has a Masters degree in physics from Caltech.