Zero-shot Sonnet Generation with Discourse-level Planning and Aesthetics Features

Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 3587 - 3597 July 10-15, 2022 ©2022 Association for Computational Linguistics Zero-shot Sonnet Generation with Discourse-level Planning and Aesthetics Features Yufei Tian Computer Science Department, University of California, Los Angeles yufeit@cs.ucla.edu Nanyun Peng Computer Science Department, University of California, Los Angeles violetpeng@cs.ucla.edu Abstract Poetry generation, and creative language gener- ation in general, usually suffers from the lack of large training data. In this paper, we present a novel framework to generate sonnets that does not require training on poems. We design a hierarchical framework which plans the poem sketch before decoding. Specifically, a con- tent planning module is trained on non-poetic texts to obtain discourse-level coherence; then a rhyme module generates rhyme words and a polishing module introduces imagery and simi- les for aesthetics purposes. Finally, we design a constrained decoding algorithm to impose the meter-and-rhyme constraint of the gener- ated sonnets. Automatic and human evalua- tion show that our multi-stage approach with- out training on poem corpora generates more coherent, poetic, and creative sonnets than sev- eral strong baselines.1 1 Introduction A sonnet is a fourteen-line poem with rigorous meter-and-rhyme constraints. In this paper, we aim at generating full-length sonnets that are logically and aesthetically coherent, without training on po- etic texts. There are several challenges for this ambitious goal. First, there are limited number of sonnets available to train a fully supervised model. The only resource is a mere 3,355 sonnets collected by Lau et al. (2018) in Project Gutenberg (Hart, 2004), one of the largest free online libraries for English literature. While it is possible to train on related corpus such as general poems or English lyrics (Ghazvininejad et al., 2016), such approaches are not applicable to many languages for which sizable poetry/lyrics data do not exist. Moreover, even if large-scale creative texts exist, learning from and mimicking existing corpora is not creative by definition and is unlikely to result in novel content. 1Our code and data are available at https://github. com/PlusLabNLP/Sonnet-Gen. a. Content Planning (trained on news and stories) Title: A Retrospect Line 1: recall, time, consider Line 2: dark, gaze, stars Line 3: day, bright, [R3] …Line 4: fog, white, [R4] Line 13: know, youth, trust Line 14: romantic, love, [R14] b. Rhyme Pairs Generation (the CMU dictionary) Line 1: Recall, time, consider Line 3: day, bright, glitter … Line 2: dark, gaze, stars Line 4: fog, white, Mars Rhyme Scheme: ABAB CDCD EFEF GG Recall the time and having me consider, x/x / /x x / x x/ d. Sonnet Generation (trained on news and stories) Meter- Constrained Decoding c. Polishing for Aesthetics (Imagery pairs, simile phrases) Imagery: Line 3: day à sun Line 14: love à rose Simile: Line 3: bright like diamond Line 4: white like snow Discourse- level Coherence Aesthetics … (/x/) (/x/x) Figure 1: An overview of our approach. The content planning module generates keywords while maintain- ing discourse-level coherence. …

Content truncated to 500 words