Norwegian poetry generation and rhyme modelling

Abstract

Computational Creativity is by some seen as the ultimate challenge for AI. Numerous works on the computationally creative task of poetry generation have been published. However, no contributions on this topic have been made for Norwegian. In this thesis, we address that gap. We present NoRSC: Norwegian Rhyme Scheme Corpus, a publicly available rhyme scheme annotated data set of Norwegian poetry. Using the NoRSC data set, we train LSTM-based models on rhyme and poetry generation. We explore combining rhyme models and language models to enhance rhyming ability in poetry generation. Our poetry generation model can generate a stanza with any rhyme scheme. Our poetry generation model is evaluated with human evaluation. We investigate whether people believe that the generated poetry is written by a human in a Turing-like test. The evaluators are also asked to score the rhyme in the generated stanzas. We received 27 answers to our web forms, and could conclude that our best poetry generation model achieved high rhyme scores, and was in about half of the instanced able to imitate a human poet.