Abstract Meaning Representations As Linked Data

The complex relationship between natural language and formal semantic representations can be investigated by the development of large, semantically-annotated corpora. The "Abstract Meaning Representation" (AMR) formulation describes the semantics of a whole sentence as a rooted, labeled graph, where nodes represent concepts/entities (such as PropBank frames and named entities) and edges represent relations between concepts (such as verb roles). AMRs have been used to annotate corpora of classic books, newstext and biomedical literature. Research on semantic parsers that generate AMRs from text is progressing rapidly. In this paper, we describe an AMR corpus as Linked Data (AMR-LD) and the techniques used to generate it (including an open-source implementation). We discuss the benefits of AMR-LD, including convenient analysis using SPARQL queries and ontology inferences enabled by embedding into the web of Linked Data, as well as the impact of semantic web representations directly derived from natural language.