You are here
DH & HLT Seminar "Abstract Meaning Representation"
Speaker: Claire Bonial, University of Colorado
Title: "Abstract Meaning Representation"
The goal of the Abstract Meaning Representation (AMR) (Banarescu et al., 2013) annotation project is to create a large-scale semantics bank using simple structures, much like the Penn Treebank (Marcus & Marcinkiewicz, 1993) has done for syntax. Such a resource would support Natural Language Processing applications such as semantic parsing, natural language generation, and machine translation. However, the question of what content belongs within a meaning representation is not a simple one. AMR to a large degree builds on PropBank (Palmer, et al., 2005) and captures many of the same concepts, such as who is doing what to whom. Consider the sentence "The team made progress even though the field was wet". PropBank annotations examine one relation at a time, starting with the verbs "make" and "be". Although previous versions of PropBank would have delved no further than the verb semantics, current practices recognize that "make progress" is a light verb construction, as well as noting the presence of an adjective predicate, "wet", after the copula. Thus, the complex predicate "make_progress" and the adjective predicate "wet" would be annotated with their arguments. Similarly, AMR creates a representation rooted by the relation progress and its arguments. However, AMR also includes discourse concepts similar to Discourse Treebank concepts (Prasad et al., 2007), which are not marked by PropBank, including the presence of a “concession” here: even though the field was wet. PropBank has served as the foundation upon which AMR has been built, and yet now many of the motivations for AMR have inspired changes underway in PropBank. Specifically, both PropBank and AMR are working towards representations that capture meaning beyond what is reflected in syntax or a predicate’s part of speech. Thus, realizations of "progress" the verb, "progress" the noun, or "make_progress" will be annotated as a unified concept. The challenges of this research, as well as the practices and goals of AMR and PropBank are discussed.