Enhancing transparency in source domain disambiguation for metaphor analysis: a cross-lingual approach integrating lexical resources, word embeddings, and human annotation

Abstract

Contemporary cognitive-linguistic research often seeks to consolidate metaphorical expressions into systematic mappings between source and target domains. However, the formulation of such mappings in natural language remains insufficiently systematized, frequently relying on intuition or on lexical resources that are not available for all languages. In this study, we propose a systematic, semi-automatic approach to source domain identification that enhances transparency, objectivity, and replicability in metaphor analysis while reducing annotator reliance on intuition. We build on an established semantic ontology, bilingual lexical resources, and distributional semantic representations to assign semantic domains to words, which serve as proxies for conceptual source domains. We manually validate the data and quantitatively evaluate the method via automatic metrics. Furthermore, we perform a qualitative evaluation of annotation disagreements and a detailed error analysis. Results indicate that the approach provides a promising foundation for semantic tagging and metaphor analysis in Slovene. The qualitative analysis of disagreements demonstrates how individual linguistic variation and cognitive biases influence domain attribution, and often prevent reaching a complete consensus between annotators. The error analysis further identifies specific limitations of the proposed approach, which arise from gaps in lexical resources and from the inherent properties of distributional semantic modeling. Overall, the findings underscore both the methodological challenges of automatic domain attribution and the cognitive complexity of source domain mapping in metaphor analysis.