Large Language Models for Metaphor Detection: Bhagavad Gita and Sermon on the Mount

Metaphors and associated literary devices were central to the composition of ancient religious and philosophical texts. Metaphors help portray spiritual messages with references to objects and situations that have deep symbolic meaning. However, the structural and contextual complexity of religious metaphors often poses a challenge in sentiment analysis. This complexity varies with different philosophical and religious traditions. There is a great need for comparative research to understand how various religious traditions are conceptualizing the elements of their experience. Recent innovations with deep learning have enabled the development of large language models (LLMs) capable of detecting metaphors. The Bhagavad Gita and the Holy Bible are central texts to Hinduism and Christianity, respectively. These texts feature a wide range of metaphors and literary devices to portray religious themes. In this paper, we use deep learning-based language models for detecting metaphors in the Bhagavad Gita and the Sermon on the Mount of the Holy Bible. We considered selected English translations of the Bhagavad Gita and Sermon on the Mount to evaluate the impact of the translation with changes in vocabulary on the detection of metaphors using LLMs. Our results show that the LLMs recognized the majority of the metaphors and the metaphorical counts in the respective translations of the religious texts. In qualitative analysis (expert review), we found that the metaphors detected have a fair consistency among translations, although the vocabulary greatly differs amongst them. Our study motivates LLMs for metaphor detection and analysis in a wide range of religious and philosophical texts.