SEEM-CZ: Annotation and Classification of Epistemic Markers in Czech

Abstract

We present a project focused on linguistic description, annotation and automatic classification of the so-called epistemic markers in Czech. These expressions, such as pravděpodobně ‘probably’, zřejmě ‘apparently’ and určitě ‘certainly’, typically operate within the pragmatic domain of language. We introduce a dataset containing manual annotations of the 40 most frequent epistemic markers in Czech, totalling almost 4,000 uses. This annotation was created using parallel InterCorp data (in Czech and English) and the TEITOK tool. We describe the annotation scheme used, the annotation process and data handling. The dataset forms the core of the emerging lexical database of these expressions (SEEMLex). Thanks to the comprehensive manual annotation, the dataset can also serve as a source of further pragmatic information and can be used as a basis for further linguistic research. The proposed annotation scheme can also be used for other languages. To demonstrate the dataset’s utility for automatic classification, we trained XLM-RoBERTa classifiers using 10-fold cross-validation, achieving 72.6% accuracy for type of use classification (6 classes) and 54.2% accuracy for degree of certainty classification (4 classes).

Resources

Details