Methodological Aspects of Semantic Annotation

This paper constitutes a preliminary report on the work carried out on semantic content annotation in the LIRICS project, in close collaboration with the activities of ISO TC 37/SC 4/TDG 3. This consists primarily of: (1) identifying commonalities in alternative approaches to the annotation and representation of various types of semantic information; and (2) developing methodological principles and concepts for identifying and characterising representational concepts for semantic content. The LIRICS project does not aim to develop a standard format for the annotation and representation of semantic content, but at providing well-defined descriptive concepts. In particular, the aim is to build an on-line registry of definitions of such concepts, called ‘data categories’, in accordance with ISO standard 12620. These semantic data categories are abstract concepts, whose use is not restricted to any particular format or representation language. We advocate the use of the metamodel as a tool to extract the most important of these abstract overarching concepts, with examples from dialogue act, temporal, reference and semantic role annotation. 1 See: http://let.uvt.nl/research/ti/iso-tdg3. 1. Models and Metamodels Alternative approaches to the marking up of linguistic resources differ most importantly in the categories of information that they aim to capture. The choices made in this respect can be represented by specifying the classes of objects and relations that are covered by their markup tags. Such a characterisation is called a model. Looking for commonalities in alternative approaches implies comparing their underlying models. This can be done by moving to a more abstract level than that of the models themselves, building a so-called metamodel. Metamodels are well known from software engineering, where they are loosely defined as a model that describes a set of models. Bunt and Romary (2004) have proposed a more formal interpretation of the term metamodel by relating it to the notion of model as used in model-theoretic semantics. This notion of metamodel can be used as a methodological tool for the definition of semantic concepts, and for the isolation of corresponding semantic data categories of importance. We argue that by using metamodels we can find an overarching conceptualisation for diverging linguistic theories. A metamodel is constructed by identifying the data categories of differing models that represent identical, similar or related items conceptually, and then by introducing a broader concept that includes the variations. In this way, one can retain the individual distinctions of the specific theories, while at the same time capturing the generalities. This is not always an easy or straightforward process, but when a metamodel is abstracted from individual models within the same theoretical area, and can also be shown to ‘fit’ the phenomena and structure to be found in the varying component theories, then this provides a good basis for consensus within the research community, as well as the first step in the standardisation of core concepts within a field. 2. Types of Semantic Annotation The LIRICS project will tackle at least the following specific areas of semantic interest: dialogue acts, temporal entities and relations, reference and semantic roles. There are very clear motivations for considering the areas discussed here in particular. Firstly, they largely coincide with similar areas of interest in ISO. Secondly, each of these areas has achieved a certain level of maturity in the semantics (and pragmatics) research communities.