A Dataset of Semantically Related Multiword Terms of the Electrical Engineering Domain

This paper presents the EEMWT dataset, a collection of 1873 triplets of co-hyponymic multiword terms of the electrical engineering domain. Each triplet combines an anchor term with closely and distantly related co-hyponymic terms. The degree of semantic relatedness is determined by the presence or absence of shared domain-specific semantic features. The primary purpose of this dataset is to serve as a tool for evaluating domain-specific language representation models and embedding vector pooling techniques by assessing the semantic relatedness of co-hyponymic multiword terms. The novelty of the dataset lies in the development approach, which replaces intuition-based scaling of relatedness with linguistically justified semantic-feature-based judgment. Furthermore, the traditional method of calculating interrater agreement rate is replaced with a statistical analysis of vector space distance between low and highly related terms. These innovations in dataset construction and evaluation have the potential to significantly reduce the costs associated with expert questionnaires in dataset development.