Kasper Aalberg Røstvold


2020

pdf
Sentimental Poetry Generation
Kasper Aalberg Røstvold | Björn Gambäck
Proceedings of the 17th International Conference on Natural Language Processing (ICON)

The paper investigates how well poetry can be generated to contain a specific sentiment, and whether readers of the poetry experience the intended sentiment. The poetry generator consists of a bi-directional Long Short-Term Memory (LSTM) model, combined with rhyme pair generation, rule-based word prediction methods, and tree search for extending generation possibilities. The LSTM network was trained on a set of English poetry written and published by users on a public website. Human judges evaluated poems generated by the system, both with a positive and negative sentiment. The results indicate that while there are some weaknesses in the system compared to other state-of-the-art solutions, it is fully capable of generating poetry with an inherent sentiment that is perceived by readers.
Search
Co-authors
Venues