Formalising Multi-layer Corpora in OWL DL - Lexicon Modelling, Querying and Consistency Control

Formalising Multi-layer Corpora in OWL DL – Lexicon Modelling, Querying and Consistency Control Aljoscha Burchardt1, Sebastian Padó2∗, Dennis Spohr3∗, Anette Frank4∗and Ulrich Heid3 1Dept. of Comp. Ling. 2Dept. of Linguistics 3Inst. for NLP 4Dept. of Comp. Ling. Saarland University Stanford University University of Stuttgart University of Heidelberg Saarbrücken, Germany Stanford, CA Stuttgart, Germany Heidelberg, Germany albu@coli.uni-sb.de pado@stanford.edu spohrds,heid@ims.uni-stuttgart.de frank@cl.uni-heidelberg.de Abstract We present a general approach to formally modelling corpora with multi-layered anno- tation, thereby inducing a lexicon model in a typed logical representation language, OWL DL. This model can be interpreted as a graph structure that offers flexible querying func- tionality beyond current XML-based query languages and powerful methods for consis- tency control. We illustrate our approach by applying it to the syntactically and semanti- cally annotated SALSA/TIGER corpus. 1 Introduction Over the years, much effort has gone into the creation of large corpora with multiple layers of linguistic an- notation, such as morphology, syntax, semantics, and discourse structure. Such corpora offer the possibility to empirically investigate the interactions between different levels of linguistic analysis. Currently, the most common use of such corpora is the acquisition of statistical models that make use of the “more shallow” levels to predict the “deeper” levels of annotation (Gildea and Jurafsky, 2002; Milt- sakaki et al., 2005). While these models fill an im- portant need for practical applications, they fall short of the general task of lexicon modelling, i.e., creat- ing an abstracted and compact representation of the corpus information that lends itself to ’linguistically informed’ usages such as human interpretation or integration with other knowledge sources (e.g., deep grammar resources or ontologies). In practice, this task faces three major problems: ∗At the time of writing, Sebastian Padó and Dennis Spohr were affiliated with Saarland University, and Anette Frank with DFKI Saarbrücken and Saarland University. Ensuring consistency. Annotation reliability and consistency are key prerequisites for the extraction of generalised linguistic knowledge. However, with the increasing complexity of annotations for ’deeper’ (in particular, semantic) linguistic analysis, it becomes more difficult to ensure that all annotation instances are consistent with the annotation scheme. Querying multiple layers of linguistic annotation. A recent survey (Lai and Bird, 2004) found that cur- rently available XML-based corpus query tools sup- port queries operating on multiple linguistic levels only in very restricted ways. Particularly problematic are intersecting hierarchies, i.e., tree-shaped analyses on multiple linguistic levels. Abstractions and application interfaces. A per- vasive problem in annotation is granularity: The gran- ularity offered by a given annotation layer may di- verge considerably from the granularity that is needed for the integration of corpus-derived data in large symbolic processing architectures or general lexical resources. This problem is multiplied when more than one layer of annotation is considered, for exam- ple in the characterisation of interface phenomena. While it may be possible to obtain coarser-grained representations procedurally by collapsing categories, such procedures are not flexibly configurable. Figure 1 illustrates these difficulties with a sentence from the SALSA/TIGER corpus (Burchardt et al., 2006), a manually annotated …

Content truncated to 500 words