Abstract

We implement an innovative strategy for metaphor identification and interpretation in texts using three different GPT OpenAI models. Metaphors are a very frequent cognitive and linguistic resource, pervasive in multiple types of human communication, singularly in discourse. Due to their conceptual and formal complexity, they are difficult to detect, classify and process. Methods of analysis have been often manual in spite of a large body of methodological proposals for computer metaphor processing. Even manual approaches pose serious challenges for highly trained annotators. Our purpose is to conduct a preliminary experiment to measure precision in identification and interpretation of metaphors in Spanish verbs analyzed in a set of corpus concordances. We implemented the method in a group of 10 polysemous Spanish verbs, in which at least one of the meanings is metaphorical. We ran three experiments with ChatGPT-4o and ChatGPT4-Turbo. We tested the precision of the models against a human annotated dataset of 1511 corpus contexts taken from the Verbario database. Results show precision between 85.18% and 88.29%, with one of the verbs achieving 94.47%. The best model is ChatGPT4-Turbo with precision between 82.61% and 94.47%. Explanations generated by the models are aligned with the identification task and show logic, complexity and specificity. The method is simple to implement and achieves high precision in comparison with traditional methods for metaphor detection and interpretation. Further research may include replicating the experiment with an expanded dataset, different data or additional and more complex tasks such as conceptual metaphor analysis.