Enhancing LLM Interpretation of Appraisals in Spanish Digital Discourse

This article examines the challenges Large Language Models (LLMs) face in interpreting evaluative language in digital discourse. These models often distort the semantics of evaluative expressions, hindering their accurate linguistic interpretation. The aim of the study is to determine whether integrating specialised knowledge improves a model's ability to correctly identify and classify evaluative meanings. Methodologically, the work draws on Martin and White's Appraisal Theory and includes an experimental evaluation of GPT-4 on a stratified corpus of posts from a social network. The analysis is conducted in two conditions — with and without external contextual knowledge — and the results are compared to expert annotations in terms of precision and recall. The findings demonstrate a substantial improvement in automatic classification: the accuracy of identifying evaluative categories increases, the range of detectable appraisal elements expands, and new patterns of meaning variation emerge. The conclusions emphasise that enriching LLMs with structured knowledge enhances the reliability of evaluative language analysis and provides deeper insight into how such meanings function in digital discourse. The proposed approach opens new avenues for improving automated methods for analysing evaluative meanings in linguistic research.