Mechanisms of transformers in metaphor detection and translation
Date
2024
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Metaphors present a unique challenge for translation due to their reliance on non-literal language, making them harder to interpret and process. Translation models often struggle with metaphors because evaluations are typically based on test sets dominated by literal language, obscuring specific issues related to metaphor translation. Furthermore, metaphors can vary in complexity, with some easier for transformer models to detect or translate, while others are more difficult, with certain metaphors appearing more frequently than others. Unlike literal language, metaphors rely heavily on contextual understanding to make sense of them, which can be problematic for models that may lack the necessary training to identify and interpret them accurately. This often leads to inaccurate or confusing translations, especially when no direct equivalent metaphor exists in the target language. Addressing this issue requires a deeper understanding of how models process metaphors, identifying which are easier or harder to translate. In this thesis, I present a method to locate metaphors that are challenging for translation utilizing a metaphor detection model followed by clustering the embeddings. This method then allows me to analyze the results, which revealed that there are multiple measurable features that exist in the source sentence that correlate with translation performance, providing valuable insight into why there is variation in how well some metaphors are translated.