Zero-Shot Trope Detection as Information Extraction and Textual Entailment

Advancements in Natural Language Processing (NLP) has allowed for machine crafted stories with excellent text fluency but which suffer from issues of coherence. The intentional use of tropes—recurring storytelling devices—may offer a solution. As a step towards incorporating tropes in story generation models, this work attempts to detect tropes in texts. This is approached through zero-shot learning, where neural language models are tasked with predicting classes unseen during training time. Zero-shot learning removes barriers of entry to NLP, especially helpful for this research area, as there is likely little crossover between storytelling experts and computer scientists. This work finds six tropes that can be detected in texts through a zero-shot means with comparable success to the previous trained trope detection method. Investigating 95 total tropes, the overall performance of zero-shot trope detection is no better than random guesses, suggesting trope detection is a more challenging task that could help prompt future zero-shot improvements.