Automatically Generating Rhythmic Verse with Neural Networks

Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, pages 168–178 Vancouver, Canada, July 30 - August 4, 2017. c⃝2017 Association for Computational Linguistics https://doi.org/10.18653/v1/P17-1016 Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, pages 168–178 Vancouver, Canada, July 30 - August 4, 2017. c⃝2017 Association for Computational Linguistics https://doi.org/10.18653/v1/P17-1016 Automatically Generating Rhythmic Verse with Neural Networks Jack Hopkins Computer Laboratory University of Cambridge jack.hopkins@me.com Douwe Kiela Facebook AI Research dkiela@fb.com Abstract We propose two novel methodologies for the automatic generation of rhythmic po- etry in a variety of forms. The first approach uses a neural language model trained on a phonetic encoding to learn an implicit representation of both the form and content of English poetry. This model can effectively learn common poetic de- vices such as rhyme, rhythm and allitera- tion. The second approach considers po- etry generation as a constraint satisfac- tion problem where a generative neural language model is tasked with learning a representation of content, and a discrimi- native weighted finite state machine con- strains it on the basis of form. By manip- ulating the constraints of the latter model, we can generate coherent poetry with arbi- trary forms and themes. A large-scale ex- trinsic evaluation demonstrated that partic- ipants consider machine-generated poems to be written by humans 54% of the time. In addition, participants rated a machine- generated poem to be the most human-like amongst all evaluated. 1 Introduction Poetry is an advanced form of linguistic commu- nication, in which a message is conveyed that sat- isfies both aesthetic and semantic constraints. As poetry is one of the most expressive forms of lan- guage, the automatic creation of texts recognis- able as poetry is difficult. In addition to requiring an understanding of many aspects of language in- cluding phonetic patterns such as rhyme, rhythm and alliteration, poetry composition also requires a deep understanding of the meaning of language. Poetry generation can be divided into two sub- tasks, namely the problem of content, which is concerned with a poem’s semantics, and the prob- lem of form, which is concerned with the aesthetic rules that a poem follows. These rules may de- scribe aspects of the literary devices used, and are usually highly prescriptive. Examples of different forms of poetry are limericks, ballads and sonnets. Limericks, for example, are characterised by their strict rhyme scheme (AABBA), their rhythm (two unstressed syllables followed by one stressed syl- lable) and their shorter third and fourth lines. Cre- ating such poetry requires not only an understand- ing of the language itself, but also of how it sounds when spoken aloud. Statistical text generation usually requires the construction of a generative language model that explicitly learns the probability of any given word given previous context. Neural language models (Schwenk and Gauvain, 2005; Bengio et al., 2006) have garnered signficant research interest for their ability to learn complex syntactic and seman- tic representations of natural language (Mikolov et al., 2010; Sutskever et al., 2014; Cho …

Content truncated to 500 words