Daejeon, July 27–31
[DH-Heading1]
Introduction
[DH- Default]
Large-scale annotated poetry data is useful for computational poetry research and ML teaching, since poetry analysis cannot rely on text alone but requires interpretive metadata. However, such resources remain scarce, as existing datasets are typically small or not fully open for reproducibility (Kao & Jurafsky 2012; Haider et al. 2020; Khan et al. 2021; Choi 2023). This gap reflects a practical challenge: affordable crowdsourced annotation is unreliable, expert annotation is reliable but costly (Haider et al. 2020).
Generative AI has recently emerged as an alternative way for data annotation (Tan et al. 2024). Prior work suggests that these models tend to perform better when substantial contextual information is available, so they may also perform reasonably well on popular poems. Such poems often have extensive online documentation, including interpretations, poetry guides, and critical essays from sources like the Poetry Foundation and Poets.org, and this additional context may help generative models produce more accurate metadata (Sun et al. 2024; Kandpal et al. 2023).
Building on these observations, we introduce a dataset of 51,356 poems with interpretations, themes, emotions, and sentiment generated by AI. Although these annotations are not human-verified, the dataset may still support new directions in computational poetry analysis and provide lightweight material for hands-on machine learning exercises in DH classrooms.
[DH-Heading1]
Dataset Description
[DH- Default]
This dataset contains 51,356 unique English poems drawn from Project Gutenberg and the Poetry Foundation, spanning classical, historical, and contemporary works. For each poem, the dataset provides AI-generated interpretations and structured metadata, including emotions, sentiment, and themes. Emotion labels follow the NRC Emotion Lexicon (Mohammad & Turney 2013), sentiment is categorized as positive, neutral, or negative, and themes are provided both as open-vocabulary descriptors and as labels drawn from the top 50 themes appearing in the Poets.org.
[DH-Heading1]
Methodology
[DH- Default]
To build this dataset, we first merged the two poetry collections, removing duplicate poems while preserving all original metadata and text formatting to avoid losing structural information essential to poetic analysis. We then used gpt-4o-2024-05-13 to generate a structured interpretation for each poem. The prompting template, available in the project repository, was designed to elicit multi-dimensional literary analysis, including thematic insights, identification of literary devices, historical or cultural context, emotional tone and mood, and structural observations related to form. The total cost for generating interpretations was approximately $2,000. In a second step, we automatically derived categorical from the generated interpretations using an on-premises gpt-oss-120b.
Statistical Summary
The generated interpretations have an average length of about 665 words and frequently refer to a range of literary devices, including imagery, symbolism, and metaphor (Figure 1). As shown in Figure 2-a, sadness is the most common primary emotion (45.2%), followed by joy (20.6%) and fear (11.7%), while other emotions appear less frequently. Sentiment labels (Figure 2-b) similarly skew toward negative sentiment (56.8%), with positive (35.8%) and neutral (7.4%) sentiment occurring less often. Detailed statistics, additional tables, and sample data are provided in the project repository, available at https://anonymous.4open.science/r/C120D891AU/.
Figure 1. Top 10 most frequently appearing literacy devices in the generated interpretations.
Figure 2. Distribution of (a) primary emotion labels and (b) sentiment labels across poems
Potential Use Cases in DH
This dataset opens several use cases in DH.
Conclusion
We introduced a large-scale, open dataset of 51,356 English poems annotated with AI-generated interpretations, themes, emotions, and sentiment. While the dataset is not a human gold standard, it provides a transparent and reproducible resource for computational poetry research and DH education.