Contrastive Retrieval Methodology for Turkish Metaphor Detection and Identification

Metaphorical expressions, as a form of figurative language, are individually limited in their use. However, when both literal and non-literal meanings are considered, they are frequently used in web content. Hence, producing a balanced dataset to learn superior representations is a challenging task, and metaphor detection suffers from a limited training dataset. To alleviate this problem, we present a retrieval-based contrastive learning approach which first identifies candidate metaphors in the input text and then detects metaphorical expressions as a claim verification task in the inherently unbalanced setting of this study. Furthermore, we adapt contrastive learning to make it easier to distinguish between the literal and figurative meanings of the same expression. For the experimental setup, we extract non-literal and literal expressions along with their meanings and sample sentences from a Turkish dictionary. In the metaphor detection subtask, performance evaluation shows that sparse and dense search variations using the Turkish-e5-Large model achieve a Recall@10 (R@10) score of 0.614. Moreover, the SimCSE-TR-Contr-Sample-Meaning model achieves the highest Recall@10 (R@10) of 0.9739 on the generated test dataset for the metaphor identification subtask. In the real-world scenario, it achieves a competitive R@10 score of 0.8684, and these results clearly demonstrate that our model can generalise to this real-world scenario.