Abstract
One dimension of modernist poetry is introducing entities in surprising contexts, such as wheelbarrow in Bob Dylan’s feel like falling in love with the first woman I meet/ putting her in a wheelbarrow. This paper considers the problem of teaching a neural language model to select poetic entities, based on local context windows. We do so by fine-tuning and evaluating language models on the poetry of American modernists, both on seen and unseen poets, and across a range of experimental designs. We also compare the performance of our poetic language model to human, professional poets. Our main finding is that, perhaps surprisingly, modernist poetry differs most from ordinary language when entities are concrete, like wheelbarrow, and while our fine-tuning strategy successfully adapts to poetic language in general, outperforming professional poets, the biggest error reduction is observed with concrete entities.
Introduction
Modernist poetry is notoriously difficult to "solve" computationally. Unlike classical poetry, which relies on predictable structures like rhyme and meter, modernist works often feel idiosyncratic or even random. Paul Valéry described the "language of poetry" as a "language within a language," where words are chosen through a painstaking process called labor limae (the labor of the file) to create specific, often jarring, effects.
The authors address a fundamental question in computational linguistics: Is the word choice of a poet actually learnable by a machine, or is it so unique to the individual that it defies generalization? Specifically, they look at how poets introduce "entities" (nouns) into their work and whether a model can predict which noun a poet used when given a set of semantically similar alternatives.
Methodology
The researchers employed a transfer learning approach using a state-of-the-art bidirectional Long Short-Term Memory (Bi-LSTM) network. The model was first pre-trained on "ordinary" language from Wikipedia to establish a baseline representation. It was then fine-tuned on a specific corpus of American modernist poetry, allowing it to adjust its expectations to the more "anti-absorptive" style of poetry—where words are meant to stand out rather than blend in.
In the evaluation task, the model was presented with a verse where a noun had been removed. It had to rank a set of candidate words retrieved from WordNet to "reconstruct" the original poem. These candidates were either synonyms or words from the same high-level semantic category.
Datasets
- PoetryDB: The primary source, containing 33,118 sentences by 32 American modernist poets.
- Wikipedia: Used for pre-training the baseline model.
- WordNet: A lexical database used to generate sets of synonyms and semantically related "distractor" words for the evaluation tasks.
Experimental Results
Fine-tuning the model on poetry led to a 34% error reduction in predicting synonyms compared to the Wikipedia-only baseline. Even when tested on poets it had never seen during training, the fine-tuned model performed better than the baseline, suggesting a shared "poetic logic" in modernism that transcends individual style.
Man vs. Machine
In a head-to-head comparison on 75 held-out examples, the fine-tuned model achieved 40% accuracy, while four professional poets averaged only 35%. This indicates that the AI successfully learned patterns of word choice that are non-trivial even for human experts.
The "Concrete" Surprise
We often associate poetry with abstract concepts like love or time. However, the study found that the fine-tuned model saw its biggest performance gains when predicting concrete entities (physical objects like "wheelbarrow" or "sea").
As illustrated in the figure above, the model was particularly adept at predicting categories such as "Matter" (e.g., clay, blood, water) and "Things" (e.g., body parts). This supports the literary theory that modernist poetry's power comes from using concrete objects in "odd pairings." Because these pairings differ so significantly from standard prose, the fine-tuning process had substantial room to improve the model's performance.
Conclusion
This research demonstrates that neural language models can be effectively adapted to the nuances of modernist poetry. The finding that the biggest error reduction occurs with concrete entities highlights the unique way modernists utilize everyday objects. While the model is proficient at selecting the right word from a list, generating poetic entities from scratch remains a significant challenge. Future work could explore the role of metaphor more deeply and investigate the relationship between nouns and their surrounding verbs and adjectives.
Summary updated