Generating Topical Poetry

Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 1183–1191, Austin, Texas, November 1-5, 2016. c⃝2016 Association for Computational Linguistics Generating Topical Poetry Marjan Ghazvininejad†, Xing Shi†, Yejin Choi‡, and Kevin Knight† †Information Sciences Institute & Computer Science Department University of Southern California {ghazvini,xingshi,knight}@isi.edu ‡Computer Science & Engineering, University of Washington yejin@cs.washington.edu Abstract We describe Hafez, a program that gener- ates any number of distinct poems on a user- supplied topic. rhyme constraints.Poems obey rhythmic and We describe the poetry- generation algorithm, give experimental data concerning its parameters, and show its gener- ality with respect to language and poetic form. 1 Introduction Automatic algorithms are starting to generate in- teresting, creative text, as evidenced by recent dis- tinguishability tests that ask whether a given story, poem, or song was written by a human or a com- puter.1 In this paper, we describe Hafez, a program that generates any number of distinct poems on a user-supplied topic. Figure 1 shows an overview of the system, which sets out these tasks: • Vocabulary. We select a specific, large vocabu- lary of words for use in our generator, and we compute stress patterns for each word. • Related words. Given a user-supplied topic, we compute a large set of related words. • Rhyme words. From the set of related words, we select pairs of rhyming words to end lines. • Finite-state acceptor (FSA). We build an FSA with a path for every conceivable sequence of vocabulary words that obeys formal rhythm constraints, with chosen rhyme words in place. • Path extraction. We select a fluent path through the FSA, using a recurrent neural network (RNN) for scoring. 1For example, in the 2016 Dartmouth test bit.ly/20WGLF3, no automatic sonnet-writing system passed indistinguishability, though ours was selected as the best of the submitted systems. Figure 1: Overview of Hafez converting a user-supplied topic word (wedding) into a four-line iambic pentameter stanza. 1183 Sections 3-7 describe how we address these tasks. After this, we show results of Hafez generating 14- line classical sonnets with rhyme scheme ABAB CDCD EFEF GG, written in iambic pentameter (ten syllables per line with alternating stress: “da-DUM da-DUM da-DUM . . . ”). We then show experiments on Hafez’s parameters and conclude by showing the generality of the approach with respect to language and poetic form. 2 Prior Work Automated poem generation has been a popular but challenging research topic (Manurung et al., 2000; Gervas, 2001; Diaz-Agudo et al., 2002; Manurung, 2003; Wong and Chun, 2008; Jiang and Zhou, 2008; Netzer et al., 2009). Recent work attempts to solve this problem by applying grammatical and seman- tic templates (Oliveira, 2009; Oliveira, 2012), or by modeling the task as statistical machine trans- lation, in which each line is a “translation” of the previous line (Zhou et al., 2009; He et al., 2012). Yan et al. (2013) proposes a method based on sum- marization techniques for poem generation, retriev- ing candidate sentences from a large corpus of po- ems …

Content truncated to 500 words