At the intersection of creative text generation and literary theory, this study explores the role of literary metaphor and its capacity to generate a range of meanings. In this regard, literary metaphor is vital to the development of any particular language. To investigate whether the inclusion of original figurative language improves textual quality, we trained an LSTM-based language model in Afrikaans. The network produces phrases containing compellingly novel figures of speech. Specifically, the emphasis falls on how AI might be utilised as a defamiliarisation technique, which disrupts expected uses of language to augment poetic expression. Providing a literary perspective on text generation, the paper raises thought-provoking questions on aesthetic value, interpretation and evaluation.
— lewe in hierdie nuwe hande waar ek algoritmies kuier [life in these new hands where I socialise algorithmically]
Although creative text generation has made notable strides in recent years [
Another important question motivating this research is whether AI can enrich language. Literary metaphor is one avenue by which to address the subject, as a formative component of linguistic creativity and a renowned resource for the generation of novel meaning [13]. Drawing on metaphor theory, this study suggests that the inclusion of figurative forms enhances the perception of originality, emotivity and memorability as well as facilitates connection. The aim is by no means to deceive through human likeness, but to engage with the exciting potential of literary metaphors to introduce new ways of using language [14]. To explore this idea, we propose an LSTM-based network for figurative language generation in Afrikaans. The proposed model produces intriguing phrases with distinctive figures of speech, such as metaphor, simile and personification.
Metaphor is integral to everyday language [15]. However, this work does not pursue the creation of conventional metaphor. Instead, we prioritise creative metaphor with artistic merit, as typically expressed in poetic discourse [16]. Compared to metaphors in everyday speech, unusual and unexpected ifgures stand out [ 17] and, in a literary context, capture readers’ attention [18]. Conducting an example evaluation of our generated output, using prevalent evaluation criteria, we argue that current frameworks are unable to do justice to poetic language. This is directly connected to the evaluation problem in creative text generation [19]. Our problematisation of NLG evaluation frameworks encourages task-specific evaluation criteria.
Finally, an important contribution lies in the choice of language. Although Multilingual Large Language Models such as GPT-4 [20] and Llama 3 [21] are able to generate text in the Afrikaans language, studies [22, 23, 24, 25] and datasets [26, 27, 28] that directly focus on Afrikaans are limited. This resonates with other low-resource languages as well. In [29, 30, 31], Masakhane shows that Natural Language Processing (NLP) research in African languages is under-represented. NLP systems are currently dominated by a handful of languages [32], and Afrikaans is one of many across the world presently unable to match their progress and sophistication. On the upside, we believe that low-resource languages ofer exciting opportunities for experimentation, collaboration and growth. The sustained invention of new metaphors is a clear indication that a language is alive [33].
Defined as “language that is more expressive and/or poetic than referential in its linguistic function” [34], figurative language comprises metaphor, metonymy, simile, personification and various other ifgures of speech. Metaphor can be defined as “the phenomenon whereby we talk and, potentially, think about something in terms of something else” [35], revealing connections between concepts that are not necessarily apparent. It is a vital resource of creative writing [36] often associated with originality. In fact, original metaphor is considered “the controlling element in all creative language” [37].
Literary metaphor is challenging to delineate given its relation to literary theory, which is vast and complex [38]. Though vital to acknowledge, it is beyond this paper’s scope to suficiently account for the range and depth of contemporary metaphor theory. For the purposes of this discussion, our emphasis falls on inventive uses of language whilst recognising the inherent dificulty in objectively diferentiating literary from ordinary language. The (dis)continuity between literary and nonliterary metaphor is subject to much debate, but there is some scholarly consensus that the former tends to exhibit more creativity, novelty, originality, complexity and interpretive dificulty [ 14]. Foregrounding its imaginative nature, Gibbs aptly describes metaphor as the “dreamwork of language” [39].
Studies in figurative text generation include simile [ 40, 41], slogan [42] and metaphor [
Our study difers in three primary ways. First, we train an LSTM-based model from scratch, without any restrictions, constraints or classifiers. We do not borrow any large-scale language models [ 46, 47] for fine-tuning, or large-scale corpora [ 48] for textual entailment. Regardless, these options are not available to low-resource languages. In this sense, our network’s advantage lies in its simplicity. Second, we approach the creative text generation process from a literary perspective, investigating ways in which literature and NLG might inform and benefit from one another. Third, we discuss the challenge of evaluating figurative language and emphasise the importance of well-defined task-specific criteria. ons biblioteek by die werkwoord gekaap die wêreld sê ek met boeke wêreldletterkunde in armoede saggies soos ’n spokerigheid in die vlug van papier sy vingers draai om haar gevoel ek het ’n gloeiende noordgrens woede is jou mond brand my in die oggendlug die wind stoppelbaard vorentoe sy kyk verras op, sy oë verlate verandering speel as foto’s van die wind die petrolbomme wat nie vertel nie sukkel is hulle kuns onbeskermde skittering in die woord my rug se wit greep ek is geld want niks kan bloei nie aarselend weerskante van die staar demokrasie was ’n daktuin begin die sonsopkoms voor die dak van my gesig jou uitgespoel is ’n onderstebo losgewoel gesprekke vir die oomblik skoongeskraap bleek
Translation (English) our library hijacked at the verb the world I say with books world literature in poverty softly like a ghostliness in the flight of paper his fingers wrap around her feeling I have a glowing northern border anger is your mouth burn me in the morning air the wind stubbles forth she looks up in surprise, his eyes deserted change plays as photos of the wind the petrol bombs that do not tell struggling is their art unprotected brightness in the word my back’s white grip I am money because nothing can bleed hesitant on either side of the stare democracy was a roof garden begins the sunrise before the roof of my face your rinsed-out is an upside-down tossed-loose conversations momentarily clean-scraped pale
We use a two-layer vanilla LSTM architecture [49], which consists of two LSTM layers with dropout layers, a fully connected layer and a softmax layer. The model was trained on a single literary novel titled Die biblioteek aan die einde van die wêreld (literally, The Library at the End of the World) [50]. The English translation is published as A Library to Flee [51]. Regarding the text’s suitability as dataset, the book abounds in vivid imagery and figurative expressions, and includes Afrikaans varieties as well as some English. Broadly, this follows the same approach as [8]. In this paper, however, we explicitly focus and expand on fully automatic text generation, centring figurative language in particular. Moreover, we investigate the vital role of original metaphor in creative writing. To clarify the process, the generation of text is open-ended (referred to as unconditional text generation); user-provided inputs, known as text prompts, are not provided. In the model, we do not enforce any specific rules, model constraints, components or use complex training schemes.
In AI research, technological innovation is understandably prized. In a creative context, however, this work serves to challenge the assumption that technical state-of-the-art equates to aesthetic value. In other words, scientific excellence does not necessarily correlate with artistic excellence. The following section describes metaphor’s creative potential to inspire new modes of expression.
Table 1 provides examples of generated phrases and sentences containing figures of speech such as metaphor, simile and personification. (Note that punctuation and capitalisation were removed in some instances.) Similar to the trained data, the network outputs unique descriptive formulations.
Given our interest in creative text generation, we believe that the success of the results is not determined by the amount of similarities shared between referents, as explained by Giles et al. [52]. Instead, we adopt Black’s interaction theory of metaphor [53]. This is relevant to the study given its emphasis on the generative function of figurative language. To clarify, because literary works (and, by implication, literary metaphors) are open to interpretation, they are “capable of generating a whole range of possible meanings. They do not so much contain meaning as produce it” [54]. In this view, meaning is not static; metaphor does not draw on pre-existing likeness but instead creates new, often surprising, likeness between concepts [55, 56].
Consequently, metaphorical language invites the reader to participate in the process of meaningmaking [57], thus facilitating connection between reader and text [58]. As regards connection, Veale [59] emphasises metaphor’s interactive dimension: it draws an engaged response from the listener/reader. Furthermore, Gibbs et al. suggest that original metaphors “communicate more emotional intensity than conventional metaphor” [60]. It follows that literary metaphor is related to not only heightened creativity but emotion as well [61, 62, 63]. As a result, we prioritise unexpected associations between disparate concepts, e.g. “democracy” as a “roof garden” (see Table 1, line 19). Through this metaphor, dissimilar domains interact, giving rise to unpredictable connections and perspectives [64]. It is challenging to measure originality in generated text [65]. However, if figurative meaning does involve a “mismatch” between domains [66], one might argue: the greater the mismatch, the greater the novelty. The efect is that of defamiliarisation, a formalist technique (referring to Russian formalism, a school of literary theory) that “makes language strange”, inviting readers to see the habitual world in fresh new ways [67]. Using AI in this sense does not replicate patterns but rather challenges convention.
Evaluation is a well-known problem in NLG that, we believe, intensifies in creative text generation. Howcroft et al. [68] draw attention to the wide and varied range of evaluation criteria in NLG, and the necessity of standardisation – a sentiment echoed by Celikyilmaz et al. [69]. As a potential counterargument, Hämäläinen and Alnajjar state [70]: if a study’s problem definition, method and evaluation are not aligned, its evaluation results inevitably lack value. Addressing the issue from a literary perspective, Van Heerden and Bas [71] propose the reconceptualisation of evaluation methods for creative systems, in particular. They argue that, though suited to standard text generation, commonly used frameworks do not encapsulate the nuances of poetic language and form.
In our study, similarly, there is a clear misalignment between our objectives and frequently used evaluation categories. To illustrate this, we apply to our results a selection of prevalent categories identified by Van der Lee et al. [ 72] – specifically fluency , coherence, accuracy and informativeness. To clarify, although these are intended for text generation in general, we chose them since there are no standardised evaluation frameworks for creative text. Applying the criteria, the generated output appears grammatically satisfied. However, it often lacks internal consistency and accuracy – as in, for instance, “I am money because nothing can bleed” (see Table 1, line 17). Referring to the same example, the sentence cannot strictly be regarded as informative (though it does communicate meaning). Although applicability certainly depends on the definition and explanation of the criteria (which are frequently absent), it is likely that this set would bear negative results in evaluation. Nevertheless, the generated text arguably draws surprising and intriguing connections between disparate domains, which we find promising in a poetic context. Analysing the theatrical production AI: When a Robot Writes a Play, Van Heerden, Duman and Bas [73] argue that the script’s distinctive qualities may be positively attributed to the model’s limitations, as opposed to its seamless performance.
Hämäläinen and Alnajjar [19] identify commonly evaluated features in creative text generation, including meaning, syntactic correctness, novelty, relevance and emotional value. These criteria could be more suitable, but we agree that further research, i.e. “evaluation of evaluation” [19], is first required, specifically to determine whether an evaluation framework could speak to literary concerns and measure a text’s potential to captivate readers. Creative text generation holds interdisciplinary appeal, and the standardisation of NLG frameworks does not necessarily preclude the development of other approaches and applications.
In response to Boden’s pivotal question “what aesthetically interesting results can computers generate, and how?” [74], this study presents an LSTM-based model for figurative language generation in Afrikaans. In general terms, our research demonstrates that a boutique language model can enhance discursive creativity. The paper discusses the importance of metaphor to creative writing, presenting one possible means of achieving enhanced emotivity, depth and originality in generated text. Moreover, ifgurative language generation is used as a point of departure to reflect on possible intersections between NLG and literary frames of reference. For example, unlike some approaches in the field, we do not expect our output to meet predetermined criteria. Instead, we shift emphasis to the competence of the evaluation framework itself. Questions that may be posed in this regard include: What kind of engagement does this kind of text invite? Does the text possess any compelling qualities, and to what extent are current evaluation criteria able to value these qualities? How could these aspects guide the reconceptualisation of evaluation in creative text generation?
This study’s confluence of technical and poetic practices served as the foundation for South Africa’s, and possibly Africa’s, first AI poetry collection, Silwerwit in die soontoe: Afrikaans se eerste KI-gedigte [Silverwhite into the Distance: Afrikaans’ First AI Poetry] [75]. Described as “a watershed moment in the Afrikaans poetry tradition” [76], the book explores the creative possibilities of language with the assistance of AI. This experimental work tests the limits of the literary, in the tradition of electronic literature [77]. The human poet interwove phrases of generated text to create verse, probing how generative AI might augment and challenge the art of poetry. The ambition was to create an original work of literature, primarily, as well as contribute to responsible NLG research.
In an attempt to establish best practice, the book’s creators acquired a writer’s permission to use his manuscript as training data. For this reason, the front cover credits the writer alongside the developer and the poet. The introduction of this poetry collection explains the distinct roles of the human poet and the AI model in the creative process. It is noteworthy that the first work of AI-generated literature in the region is an instance of co-creativity that privileges the human author. This approach poses an ethical, albeit modest, alternative to LLMs in creative contexts.
Results incorporated in this paper have received funding from the European Union’s Horizon 2020 research and innovation programme under the Marie Skłodowska-Curie grant agreement No 900025. This work was also supported by the 2232 International Fellowship for Outstanding Researchers Program of TÜBİTAK [project number 18C285]; and the Engineering and Physical Sciences Research Council [grant number EP/Y009800/1].
The authors are grateful to Etienne van Heerden for sharing the original manuscript of his novel. [7] D. Uthus, M. Voitovich, R. Mical, Augmenting poetry composition with Verse by Verse, in: Proc
NAACL Industry Track, 2022, pp. 18–26. [8] I. Van Heerden, A. Bas, AfriKI: Machine-in-the-loop Afrikaans poetry generation, in: Proc EACL Workshop on Bridging Human–Computer Interaction and Natural Language Processing, 2021, pp. 74–80. [9] K. Yang, D. Klein, FUDGE: Controlled text generation with future discriminators, in: Proc NAACL, 2021, pp. 3511–3535. [10] A. Bas, M. O. Topal, C. Duman, I. Van Heerden, A brief history of deep learning-based text generation, in: Proc ICCA, 2022, pp. 1–4. [11] OpenAI, Introducing ChatGPT, 2022. https://openai.com/index/chatgpt/. [12] L. Edwards, I. Szpotakowski, G. Cifrodelli, J. Sangaré, J. Stewart, Private ordering and generative
AI: What can we learn from model terms and conditions?, CREATe Working Paper Series, 2024. [13] L. Hidalgo-Downing, Metaphor and metonymy, in: The Routledge Handbook of Language and
Creativity, Routledge, 2019, pp. 107–128. [14] E. Semino, G. Steen, Metaphor in literature, in: The Cambridge Handbook of Metaphor and
Thought, Cambridge University Press, 2008, pp. 232–246. [15] G. Lakof, M. Johnson, Conceptual metaphor in everyday language, The Journal of Philosophy 77 (1980) 453–486. [16] M. Caracciolo, Creative metaphor in literature, in: The Routledge Handbook of Metaphor and
Language, Routledge, 2016, pp. 224–236. [17] Z. Kovecses, Metaphor: A Practical Introduction, Oxford University Press, 2010. [18] G. Steen, Understanding Metaphor in Literature: An Empirical Approach, Longman, 1994. [19] M. Hämäläinen, K. Alnajjar, Human evaluation of creative NLG systems: An interdisciplinary survey on recent papers, in: Proc ACL-IJCNLP Workshop on Natural Language Generation, Evaluation, and Metrics, 2021, pp. 84–95. [20] OpenAI, Gpt-4 technical report, arXiv preprint arXiv:2303.08774 (2024). [21] Meta AI, Introducing Meta Llama 3: The most capable openly available LLM to date, 2024.
https://ai.meta.com/blog/meta-llama-3/. [22] M. Van Zaanen, G. Van Huyssteen, Improving a spelling checker for Afrikaans, in: Proc CLIN, 2003, pp. 143–156. [23] L. Sanby, I. Todd, M. C. Keet, Comparing the template-based approach to GF: The case of Afrikaans, in: Proc WebNLG, 2016, pp. 50–53. [24] P. Ziering, L. Van der Plas, Towards unsupervised and language-independent compound splitting using inflectional morphological transformations, in: Proc NAACL, 2016, pp. 644–653. [25] P. Dirix, L. Augustinus, D. Van Niekerk, F. Van Eynde, Universal dependencies for Afrikaans, in:
Proc NoDaLiDa, 2017, pp. 38–47. [26] R. Eiselen, M. Puttkammer, Developing text resources for ten South African languages, in: Proc
LREC, 2014, pp. 3698–3703. [27] L. Augustinus, P. Dirix, D. Van Niekerk, et al., AfriBooms: An online treebank for Afrikaans, in:
Proc LREC, 2016, pp. 677–682. [28] J. Roux, South African National Centre for Digital Language Resources, in: Proc LREC, 2016, pp.
2467–2470. [29] ∀, et al., Participatory research for low-resourced machine translation: A case study in African languages, in: Proc EMNLP, 2020, pp. 2144–2160. [30] ∀, et al., Masakhane–machine translation for Africa, in: Proc ICLR Workshop on AfricaNLP, 2020. [31] D. I. Adelani, J. Abbott, G. Neubig, et al., MasakhaNER: Named entity recognition for African languages, Transactions of the Association for Computational Linguistics 9 (2021) 1116–1131. [32] P. Joshi, S. Santy, A. Budhiraja, K. Bali, M. Choudhury, The state and fate of linguistic diversity and inclusion in the NLP world, in: Proc ACL, 2020, pp. 6282–6293. [33] R. L. Trask, Language: The Basics, Routledge, 2004. [34] D. Chandler, R. Munday, A Dictionary of Media and Communication, Oxford University Press, 2011. [35] E. Semino, Metaphor in Discourse, Cambridge University Press, 2008. [36] C. Baldick, The Concise Oxford Dictionary of Literary Terms, Oxford University Press, 1996. [37] P. Newmark, A Textbook of Translation, Prentice Hall, 1988. [38] M. H. Freeman, The role of metaphor in poetic iconicity, in: Beyond Cognitive Metaphor Theory:
Perspectives on Literary Metaphor, Routledge, 2011, pp. 158–174. [39] R. W. Gibbs, The process of understanding literary metaphor, Journal of Literary Semantics 19 (1990) 65–79. [40] S. Harmon, Figure8: A novel system for generating and evaluating figurative language, in: Proc
ICCC, 2015, pp. 71–77. [41] T. Chakrabarty, S. Muresan, N. Peng, Generating similes effortlessly like a pro: A style transfer approach for simile generation, in: Proc EMNLP, 2020, pp. 6455–6469. [42] K. Alnajjar, H. Kundi, H. Toivonen, “Talent, skill and support.” A method for automatic creation of slogans, in: Proc ICCC, 2018, pp. 88–95. [43] K. I. Gero, L. B. Chilton, Metaphoria: An algorithmic companion for metaphor creation, in: Proc
CHI, 2019, pp. 1–12. [44] J. Brooks, A. Youssef, Discriminative pattern mining for natural language metaphor generation, in: Proc Big Data, 2020, pp. 4276–4283. [45] I. A. Richards, The Philosophy of Rhetoric, Oxford University Press, 1936. [46] J. Devlin, M.-W. Chang, K. Lee, K. Toutanova, BERT: Pre-training of deep bidirectional transformers for language understanding, in: Proc NAACL-HLT, 2019, pp. 4171–4186. [47] M. Lewis, Y. Liu, N. Goyal, et al., BART: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension, in: Proc ACL, 2020, pp. 7871–7880. [48] A. Williams, N. Nangia, S. Bowman, A broad-coverage challenge corpus for sentence understanding through inference, in: Proc NAACL-HLT, 2018, pp. 1112–1122. [49] S. Hochreiter, J. Schmidhuber, Long short-term memory, Neural Computation 9 (1997) 1735–1780. [50] E. Van Heerden, Die biblioteek aan die einde van die wêreld, NB-Uitgewers, 2019. [51] E. Van Heerden, A Library to Flee, Tafelberg Publishers, 2022. [52] A. Giles, J. F. Kess, C. Uda, Metaphor as the creative origin of lexical ambiguity in English and
Japanese, Working Papers of the Linguistics Circle 10 (1991) 49–62. [53] M. Black, Models and Metaphors: Studies in Language and Philosophy, Cornell University Press, 1962. [54] T. Eagleton, How to Read Literature, Yale University Press, 2014. [55] B. Indurkhya, Metaphor and Cognition: An Interactionist Approach, Kluwer Academic, 1992. [56] T. Veale, E. Shutova, B. B. Klebanov, Metaphor: A Computational Perspective, Morgan & Claypool
Publishers, 2016. [57] R. M. White, The Structure of Metaphor: The Way the Language of Metaphor Works, Wiley, 1996. [58] T. Cohen, Metaphor and the cultivation of intimacy, Critical Inquiry 5 (1978) 3–12. [59] T. Veale, Metaphor in the age of mechanical production, in: Metaphor and Metonymy in the Digital Age: Theory and Methods for Building Repositories of Figurative Language, John Benjamins Publishing Company, 2019, pp. 75–98. [60] R. W. Gibbs, J. S. Leggitt, E. A. Turner, What’s special about figurative language in emotional communication, The Verbal Communication of Emotions: Interdisciplinary Perspectives (2002) 125–149. [61] L. Fainsilber, A. Ortony, Metaphorical uses of language in the expression of emotions, Metaphor and Symbol 2 (1987) 239–250. [62] T. I. Lubart, I. Getz, Emotion, metaphor, and the creative process, Creativity Research Journal 10 (1997) 285–301. [63] S. R. Fussell, M. M. Moss, Figurative language in emotional communication, Social and Cognitive
Approaches to Interpersonal Communication (1998) 113–141. [64] E. C. Way, Knowledge Representation and Metaphor, Springer Science & Business Media, 1991. [65] B. B. Klebanov, N. Madnani, Automated evaluation of writing – 50 years and counting, in: Proc
ACL, 2020, pp. 7796–7810. [66] R. J. Fogelin, Figuratively Speaking: Revised Edition, Oxford University Press, 2011. [67] V. Shklovsky, Art as technique, in: Literary Theory: An Anthology, John Wiley & Sons, 1917, pp.
15–21. [68] D. M. Howcroft, A. Belz, M.-A. Clinciu, et al., Twenty years of confusion in human evaluation:
NLG needs evaluation sheets and standardised definitions, in: Proc INLG, 2020, pp. 169–182. [69] A. Celikyilmaz, E. Clark, J. Gao, Evaluation of text generation: A survey, arXiv preprint arXiv:2006.14799 (2020). [70] M. Hämäläinen, K. Alnajjar, The great misalignment problem in human evaluation of NLP methods, in: Proc EACL Workshop on Human Evaluation of NLP Systems, 2021, pp. 69–74. [71] I. Van Heerden, A. Bas, AI as Author – Bridging the Gap Between Machine Learning and Literary
Theory, Journal of Artificial Intelligence Research 71 (2021) 175–189. [72] C. Van der Lee, A. Gatt, E. Van Miltenburg, E. Krahmer, Human evaluation of automatically generated text: Current trends and best practice guidelines, Computer Speech & Language 67 (2021) 101151. [73] I. Van Heerden, C. Duman, A. Bas, Performing the Post-Anthropocene: AI: When a Robot Writes a Play, TDR: The Drama Review 67 (2023) 104–120. [74] M. A. Boden, Creativity and Art: Three Roads to Surprise, Oxford University Press, 2012. [75] I. van Heerden, A. Bas, E. Van Heerden, Silwerwit in die soontoe: Afrikaans se eerste KI-gedigte,
Naledi, 2023. [76] N. Bennett, Die wonder van willekeur, in: Rapport, 2023, p. 14. [77] N. K. Hayles, Electronic literature: What is it, Doing Digital Humanities: Practice, Training, Research (2016) 197–226.