Scientific innovations and societal developments happen in rapid succession: quantum computers, AI, the pandemic… When communicating about such developments with non-expert audiences, journalists often use metaphors. Metaphors describe abstract, unfamiliar concepts in terms of more concrete, more accessible concepts (e.g. AI as a friend). The metaphors we choose influence perceptions and decision-making, making it crucial to study their use in media texts.
However, manual metaphor identification is very labor-intensive. On average, one in six words in news texts is used metaphorically. However, most metaphors will not be used to describe the topic of interest (e.g. AI). To speed up this process, various methods relying on supervised machine learning have been proposed. While these can achieve acceptable performance on data from the same source as the training data, performance tends to drop on data covering new topics. It is unlikely that a supervised system could identify metaphors it has never seen in its training data. Even recent models (e.g., ChatGPT), perform poorly.
To resolve this issue, we develop a hybrid computationally-assisted approach to metaphor identification, combining simple and transparent NLP methods with targeted manual annotation. The goal is to limit manual labor by identifying passages about the target topic containing potential metaphors automatically. First, we identify passages about our target topics semi-automatically. Then, we identify metaphors in these passages (initially based on manual analysis). Text passages and metaphors identified as such will subsequently be fed back into the system, creating a loop to find more relevant passages and metaphors. This approach combines machine learning with human feedback and verification procedures to combat blind spots. We aim to develop a standardized procedure applicable to various domains. As use-cases, we focus on news about covid, climate technologies and AI.
Supervisors:
- Gudrun Reijnierse
- Pia Sommerauer
Students:
- Bastiaan Sizoo
- Urtė Jakubauskaitė