A methodology for designing semantic annotations

This paper presents a methodology for designing languages for semantic annotation. Central in this methodology is the specification of representation formats as renderings of conceptual structures defined by an abstract syntax as set-theoretic constructs. An ideal representation format is defined as one that is able to represent all the conceptual distinctions made in the abstract syntax, and of which each representation encodes one and only one structure defined by the abstract syntax. The semantics of an annotation language is defined for its abstract syntax and is shared by all its representation formats; every ideal representation format is therefore convertible through a meaning-preserving mapping to any other ideal representation format. The methodology is called CASCADES after its four stages: Conceptual analysis, Abstract syntax, Semantics and Concrete syntax for Annotation DESign. The CASCADES model derives its usefulness from supporting a systematic design process for semantic annotations, giving due attention to the conceptual and semantic issues underlying choices in annotation formats, including support in the form of procedures for how to move from one stage of the design process to another, in particular for how to construct an abstract syntax given a conceptual analysis; how to define a semantics for a given abstract syntax; and how to map an abstract syntax to an XML-based concrete syntax. Three applications of the CASCADES methodology are discussed: (1) the design of an ISO standard for dialogue annotation starting from a conceptual analysis; (2) the analysis of existing annotation schemes such as those of the Penn Discourse Treebank and TimeML, as a basis for the development of ISO standards for semantic annotation; (3) the detection and repair of deficiencies in existing annotation schemes.