bert’s Interpretation of Literalmente ‘Literally’: What Deep Learning Models Can Tell Us about Synchronic Layering and Diachronic Shifts

Abstract How do language models disambiguate semantically and pragmatically complex and polysemic meanings? In this study, we present a computational approach to the analysis of Spanish’s polysemic literalmente ‘literally,’ an adverb whose meaning and pragmatic functions range from strict word-by-word denotation to (inter)subjective intensification and emphasis. Focusing on the Spanish pre-trained bert model – beto –, two objectives are pursued: i) to shed light onto how artificial language processors interpret pragmatically polyfunctional and semantically polysemic words, and ii) to showcase how the contextual cues drawn on by an artificial language processor can help elucidate semantic polysemy and change in natural language. Using Local Interpretable Model-Agnostic Explanations ( lime ), our results show that more innovative and grammaticalized uses exhibit a higher degree of syntactic polyfunctionality. We discuss parallelisms and cross-pollination potential between the uncovered computational dynamics of polysemic literalmente and theories of grammaticalization and semantic change.