Computational Linguistic Creativity: Poetry generation given visual input
Abstract
In this Thesis, a system capable of generating poetry from images was designed and implemented. The implementation was then used to generate 153 stanzas. The system includes a CNN for object classification, a module to find related words and rhyme pairs and an LSTM network to score a tree search where nodes are representing the words being generated.
Three different experiments were conducted to evaluate the performance of the system. The first one looked at how training song lyrics on the LSTM network affected the word perplexity of different architectures. The second and third experiments were conducted with volunteering participants to evaluate the generated poetry.