Electronic Thesis/Dissertation
 

Automated Generation and Detection of Novel On-Topic Metaphors Using Machine Learning

Open Access

The primary goal of this research is to automatically generate novel metaphors, where a generated metaphor can be interpreted by humans and based on a topic provided as input to the generator. This research is motivated by the idea that metaphors can inspire people with fresh, creative thoughts, and foster greater trust between humans and machines by making the machines more creative and human-like in their use of language. Metaphor generation requires metaphor detection to ensure that a metaphor was generated. On-topic metaphor generation requires identifying the topic in the generated metaphor and then measuring the similarity of the identified topic to the intended topic, or measuring the relatedness of the generated text to the topic. Novel metaphor generation requires tests of originality and style-adherence. This research also introduces metrics for novelty and originality, where no standards currently exist. Our approach to metaphor generation can be organized into two primary parts. First, train a language model on the language of metaphors. Second, build quality assurance mechanisms to automatically encourage style-adherence, fluency, originality, and topic-relatedness in the outputs. We emulate the processes that humans go through when they are trying to think of a metaphor to include in their creative writing. We may generate many candidates that fail our quality assurance tests before generating one that we would put in writing. To assess the subjective quality of our generated metaphors, we conducted two experiments with human participants. We found that our trained models performed better than random on the task of generating a metaphor on-topic, and there was a positive correlation (0.625) between whether the topic was identifiable and a participant thought that the metaphor was written by a human. Furthermore, we found that the participants, on average, thought that 60% of the randomly selected metaphor generations were written by a human and inspiring and rich in meaning.

Author Language Keyword Date created Type of Work License
  • All rights reserved
Rights statement GW Unit Degree Advisor Committee Member(s) Persistent URL

Notice to Authors

If you are the author of this work and you have any questions about the information on this page, please use the Contact form to get in touch with us.

Thumbnail Title Date Uploaded Visibility Actions
Preview of Brooks_gwu_0075A_15883.pdf Brooks_gwu_0075A_15883.pdf 2022-04-25 Open Access