Tools & methods
How the eighty-eight contexts were produced, and the software that produced them. The reception is not asserted but computed, read and adversarially checked; this page names each tool and how it was used, so the work can be scrutinised and re-run.
Each context is the product of four passes, and every pass is grounded, not invented.
- Discovery. A TF-IDF cosine baseline (scikit-learn) and
BGE-M3 dense and sparse embeddings (FlagEmbedding) locate the poems that
echo the Canção and cluster those that echo one another; verbatim
n-gram matching extracts the reused fragments that become the
INT1passage anchors. - Interpretation. A Large-language-model reads each candidate against the
controlled
INT3relation vocabulary and itsINT2motifs, and proposes the equivalence-and-difference rationale for human review. - Refutation. A further adversarial pass asks, of every candidate, whether the link is real at all. Weak or oblique cases are held back rather than asserted (three were refuted, fifteen quarantined for human review).
- Grounding. The relation type, the motifs and the modes are each bound to a real SKOS concept in a shared vocabulary, so that a ratified context promotes into the live RPPA graph unchanged.
Alongside these, a structural signal is computed independently: spaCy with the Portuguese model measures the T0–T6 syntactic-template degree of reuse of “Minha terra tem ___”, after Enslen & Bell (2024). Enslen’s own similarity coefficient and Positive / Negative / Other mode (from Song of Exile, parsed out of the book’s appendix) are carried as a cross-check, never copied.
The core analytical toolchain, with the versions used for this study. TextPAIR is a potential method held in reserve for larger corpora; how the interpretive model is used and checked is detailed in the next section.
| Tool | Role in this study | Reference |
|---|---|---|
| BGE-M3 via FlagEmbedding 1.3.5 | Multilingual dense & sparse embeddings for discovery and reuse-clustering (BAAI/bge-m3). |
Chen et al. 2024, arXiv:2402.03216 |
| scikit-learn 1.x | TF-IDF vectorisation and cosine-similarity discovery baseline. | Pedregosa et al. 2011, JMLR 12 |
| TextPAIR potential | Sequence-alignment text-reuse detection at corpus scale. Evaluated as a discovery method and held in reserve for scaling beyond the eighty-eight (see Limitations). | ARTFL Project, Univ. of Chicago |
| Claude Opus 4.8 (Anthropic) | Relation / motif classification, the equivalence-and-difference proposal and adversarial refutation. | Anthropic, Claude |
| spaCy 3.8 + pt_core_news_md | Portuguese parsing for the T0–T6 syntactic-template signal. | Honnibal et al. 2020, Zenodo |
| NumPy 1.26 | Vector arithmetic underlying the similarity measures. | Harris et al. 2020, Nature 585 |
| supporting PyTorch, Transformers, sentence-transformers | The deep-learning stack the embedding model runs on. | Paszke 2019; Wolf 2020; Reimers & Gurevych 2019 |
The interpretive layer, and how it is checked
The relation typing, the motifs and the equivalence-and-difference proposals were produced by computational means, not by a human reading each of the eighty-eight poems. This is stated plainly because it matters: the proposals are candidates. They are made legible (every one carries its evidence, its confidence and an adversarial verdict), demarcated as unratified, and offered for expert peer-review. A ratified proposal is a state-change, not a rebuild. The candidate layer thus decouples generation from verification: nothing here claims to be settled scholarship until a domain expert has ratified it.
Every relation type, motif and mode resolves to a concept in a shared vocabulary rather than free text: the relation typology maps to the Princeton Encyclopedia of Poetry & Poetics, Burwick’s Romanticism: Keywords and the Eighteenth-Century Poetry Archive (ECPA); the symbols to Ferber’s Dictionary of Literary Symbols; the modes and devices to the Princeton Encyclopedia; and the poem’s own structural motifs (the “Minha terra tem” template, terra, saudade) to a small study-specific scheme drawn from Enslen’s Song of Exile (2022) and Enslen & Bell (2024). Full bibliographic detail is on the overview and in the references below.
- The candidates are machine-generated and await expert ratification (above).
- Geography is country-level. Within-country marker placement on the map is schematic (spread across the period’s literary centres), not a claim of exact locality.
- Corpus-scale text-reuse tooling was evaluated, not used. TextPAIR was tested and set aside, because its frequency filtering would suppress “minha terra tem”, the very signal of interest, at this corpus size; verbatim n-gram matching was used instead.
- Eighty-eight is a deep sample, not the whole reception. Enslen catalogues some five hundred texts; this study goes deep on the nineteenth-century core.
The study is data-driven: the pages render a standalone JSON bundle, so re-running the pipeline (discovery → interpretation → grounding) regenerates the bundle and the site reflects the new reality on reload. The views themselves are drawn with standard open-source libraries (Cytoscape.js for the network, Leaflet for the map); they display the results but produce none of them.
Chen, J., Xiao, S., Zhang, P., Luo, K., Lian, D. & Liu, Z. (2024). BGE M3-Embedding. arXiv:2402.03216. · Honnibal, M., Montani, I., Van Landeghem, S. & Boyd, A. (2020). spaCy: Industrial-strength Natural Language Processing in Python. Zenodo, doi:10.5281/zenodo.1212303. · Pedregosa, F. et al. (2011). Scikit-learn: Machine Learning in Python. Journal of Machine Learning Research 12: 2825–30. · Harris, C. R. et al. (2020). Array programming with NumPy. Nature 585: 357–62. · Wolf, T. et al. (2020). Transformers: State-of-the-Art Natural Language Processing. EMNLP: System Demonstrations. · Reimers, N. & Gurevych, I. (2019). Sentence-BERT. EMNLP-IJCNLP. · Paszke, A. et al. (2019). PyTorch. NeurIPS. · Franz, M. et al. (2016). Cytoscape.js. Bioinformatics 32(2): 309–11. · Anthropic. Claude Opus 4.8 (large language model). anthropic.com. · Enslen, J. A. (2022). Song of Exile. Purdue University Press. · Enslen, J. A. & Bell, J. R. (2024). Minha Terra Tem ____. MATLIT 10(1): 107–24. · Ferber, M. (2017). A Dictionary of Literary Symbols (3rd ed.). Cambridge University Press. · The Princeton Encyclopedia of Poetry and Poetics (4th ed., 2012). Princeton University Press. · Burwick, F. (2015). Romanticism: Keywords. Wiley-Blackwell.