Summary
Metaphors are important for how humans think and communicate. The authors study how large language models (LLMs) understand and use metaphors by identifying specific parts of the models that control them. They show that by adjusting these identified parts, they can change how metaphorical the models’ responses are, both in classifying language and creating new text. This means it might be possible to control how figurative the AI's language is.
What this means in practice
- •For chatbot developers: Adjust chatbots to vary the use of metaphors for different conversational tones or styles.
- •For digital content creators: Generate text with controllable metaphor levels to fit different creative writing or marketing needs.
Authors
Shihui Li, Mette Andersen, Martha Lewis
Abstract
Metaphor is a crucial aspect of human cognition. As such, we would like to understand how it is represented in large language models (LLMs) and how it might be controlled. We use mechanistic interpretability techniques to identify causal subspaces in a range of LLMs covering different architectures and sizes. We show that we are able to steer output on classification tasks across different parts of speech. We are also able to steer output on generation tasks, indicating that these techniques can be used to control the metaphoricity of LLM output.