PINEAPPLE: Personifying INanimate Entities by Acquiring Parallel Personification Data for Learning Enhanced Generation

Proceedings of the 29th International Conference on Computational Linguistics, pages 6270–6284 October 12–17, 2022. 6270 PINEAPPLE: Personifying INanimate Entities by Acquiring Parallel Personification Data for Learning Enhanced Generation Sedrick Scott Keh 1, Kevin Lu 2, Varun Gangal∗1, Steven Y. Feng∗3, Harsh Jhamtani 1, Malihe Alikhani 4, Eduard Hovy 1 1Carnegie Mellon University, 2University of Waterloo, 3Stanford University, 4University of Pittsburgh {skeh,vgangal,jharsh,hovy}@cs.cmu.edu, syfeng@stanford.edu kevin.lu1@uwaterloo.ca, malihe@pitt.edu Abstract A personification is a figure of speech that en- dows inanimate entities with properties and actions typically seen as requiring animacy. In this paper, we explore the task of person- ification generation. To this end, we pro- pose PINEAPPLE: Personifying INanimate Entities by Acquiring Parallel Personification Data for Learning Enhanced Generation. We curate a corpus of personifications called Per- sonifCorp, together with automatically gener- ated de-personified literalizations of these per- sonifications. We demonstrate the usefulness of this parallel corpus by training a seq2seq model to personify a given literal input. Both automatic and human evaluations show that fine-tuning with PersonifCorp leads to signif- icant gains in personification-related qualities such as animacy and interestingness. A de- tailed qualitative analysis also highlights key strengths and imperfections of PINEAPPLE over baselines, demonstrating a strong ability to generate diverse and creative personifications that enhance the overall appeal of a sentence. 1 1 Introduction Personification is the attribution of animate actions or characteristics to an entity that is inherently inan- imate. Consider, for example, the sentence “The stars danced playfully in the moonlit sky.” Here, the vibrance of the stars (something inanimate) is being likened to dancing playfully, which is a distinctly animate action. By allowing readers to construct clearer mental images, personifications enhance the creativity of a piece of text (Bloom- field, 1980; Dorst, 2011; Flannery, 2016). Being able to automatically identify and generate personifications is important for multiple reasons. First, humans naturally use personifications when communicating. When we say something like “My ∗ Equal contribution by Varun and Steven 1Data and code can be found at https://github. com/sedrickkeh/PINEAPPLE Figure 1: Overall PINEAPPLE model pipeline. The left part of the diagram shows the corpus creation pro- cess, while the right part of the diagram shows the train- ing and generation process. phone has died,” or “My car is not cooperating,” to a dialogue system, it is important that the dialogue system understands the intended meaning behind these personifications. If these systems interpret personifications literally, they may fail in several downstream tasks (e.g. classification) since their understanding is incorrect. Being able to generate personifications also allows dialogue agents and language models to be more creative and generate more figurative sentences. Personification genera- tion has additional applications such as AI-assisted creative writing, since machine-generated figures of speech have been shown to enhance the interest- ingness of written text (Chakrabarty et al., 2021). Despite previous success in generating other fig- ures of speech such as similes (Chakrabarty et al., 2020), metaphors (Stowe et al., 2021), hyperboles (Troiano et al., 2018), irony (Van Hee et al., 2018), and sarcasm (Hazarika et al., 2018; …

Content truncated to 500 words