RunoVerse

Poem & Verse Similarity Guide

What is this page?

A reference guide to the similarity system in RunoVerse. It documents all similarity methods (five poem-level algorithms plus two fused rankings, and five verse-level algorithms plus a combined ranking), the geography metrics, and the tools that use them to find related poems and verses across 292K poems and 4.3M verse lines.

How to navigate

Each section describes one tool — what it does, its key features, and a direct link to open it. The tables at the bottom explain the poem-level algorithms, the verse-level algorithms, and the geography metrics. Scroll or use the browser's find-in-page (Ctrl/Cmd+F) to jump to a section.

Related explorer pages

Similarity Explorer — poem and verse similarity search · Poem Reader — per-poem similarity tabs · Verse Network — interactive similarity graph · Formulaicity — per-verse formula scores and the By Formula (clusters) browser · Path Finder — shortest verse chains · Verse Analysis — cross-algorithm dashboard

RunoVerse provides multiple ways to explore how poems and verses relate to each other across the Finnic runosong traditions. Whether you are tracing the spread of a formulaic verse line, comparing thematically related poems from different regions, or mapping cross-lingual parallels, the tools below offer complementary perspectives on the roughly 292,000-poem, 4.3-million-verse corpus.

Poem Reader

The Poem Reader is an interactive viewer for about 292,000 poems from the SKVR (Finnish, published), JR (Finnish, unpublished), KR (Finnish, published Kalevala-metre poetry) and ERAB (Estonian) collections. Each word in a poem is annotated with its standard orthography, English gloss, part-of-speech tag, and a link to its lemma in the dictionary. Clicking a word opens a dictionary lookup panel with full definitions from up to nine lexicographic sources.

Open Poem Reader →

Similarity Explorer

The Similarity Explorer is a standalone tool for investigating poem and verse similarity in depth. It provides seven views, each showing a different angle on the corpus. Press ? on the Similarity page for quick reference.

Poem Similarity lets you select any poem and view its nearest neighbors across all seven ranked lists — five poem-level algorithms plus the Verse Match and Combined (RRF) fused rankings. Each view displays its top matches with numeric similarity scores. A side-by-side comparison panel lets you read two poems together, line by line with shared words highlighted. A network graph visualizes the similarity neighborhood. Geographic and temporal analytics reveal where and when similar poems were collected.

Verse Similarity lets you search for individual verse lines and see similar verses from across the corpus, with results from five verse-level algorithms plus a combined view. An experimental Semantic neighbours – native embedding panel below each verse’s results finds lines with a similar meaning in any language directly from the original texts — including Estonian–Finnish pairs that share no words. It shows two multilingual embedding models together, each labelled — the current LaBSE model and the previous mpnet model; it is separate from the five methods, highlights lines none of the five methods found, and opens by itself when it contains such lines (otherwise it stays collapsed), with a caveat explaining its limits. A research-view control can also reveal the candidate lines its quality rule rejected, each labelled with the reason it was hidden.

Verse Pairs browses the strongest verse-to-verse connections found across the whole corpus, with a By Place Pair view grouping pairs by collection location.

Geography offers five map-based submodes: Verse Network (place-to-place connections), Estonian↔Finnish Formulas (cross-corpus bridges), Poem Connections, Combined (verse + poem layers), and Place Focus (single-place exploration).

Formulaicity reveals the formulaic skeleton of individual poems, scoring each verse by how closely it matches verses elsewhere in the corpus. Its By Formula (clusters) view browses the 200 largest formula clusters, sortable by member count, number of places, text, or cluster ID.

Compare puts any two poems side by side, verse by verse, with shared vocabulary highlighted in layers (same wordform, same lemma, same translation) and verse-to-verse connection lines.

Cross-corpus clusters browses 19,423 groups of near-identical verse lines that span the Estonian and the Finnish-side collections, with language-mix and size filters.

Open Similarity Explorer →

Verse Concordance

The Verse Concordance provides full-text search across 2.9 million unique verse lines drawn from 4.32 million total verse lines. Enter any text fragment to find matching verses. Results include a geographic distribution chart showing where each verse was collected (with a link to the full map view), a language breakdown (Estonian vs. Finnish), and occurrence tables listing which poems contain the verse and how frequently it appears.

Open Verse Concordance →

Verse Network

The Verse Network displays an interactive force-directed graph of verse similarity neighborhoods. Start from any verse and explore multi-hop connections to see how verses are linked through shared similarity across the corpus. Each edge shows a per-algorithm score breakdown, so you can see which similarity measures contribute to each connection. Results can be exported as CSV for further analysis.

Open Verse Network →

Verse Path Finder

The Verse Path Finder locates the shortest chain of similar verses connecting any two verses in the Finnic runosong corpus. The result is displayed as an interactive chain showing each intermediate verse and the similarity scores between consecutive steps. This reveals how seemingly unrelated verses may be connected through a sequence of incremental textual similarities.

Open Verse Path Finder →

Formulaicity

The Formulaicity view scores each verse of a poem by how closely it matches verses elsewhere in the corpus, from widely shared formulas to lines found nowhere else. A “formula” here is a cluster of similar verse lines found across many poems — evidence of oral tradition transmission, where singers in different times and places used recognizably similar wording. Its By Formula (clusters) view browses the 200 largest formula clusters — sortable by member count, number of places, text, or cluster ID — each showing its variant texts, member count across poems, geographic spread across collection places, and cross-links for further exploration.

Open Formulaicity →

Verse Similarity Analysis

The Verse Similarity Analysis is a cross-algorithm dashboard showing how the five verse-level similarity algorithms — Jaccard, TF-IDF, Translation-pivot, CharBigram, and Sentence — compare across the 4.3 million verse lines in the corpus. It includes formulaic cluster analysis and geographic coverage metrics, providing a high-level view of how similarity patterns distribute across the material.

Open Verse Similarity Analysis →

Poem-level similarity algorithms

Seven ranked lists identify related poems: five poem-level algorithms plus two fused rankings (verse match and combined RRF). Each captures a different aspect of similarity — from shared vocabulary to cross-lingual thematic overlap to structural verse alignment. Results from all seven are available in the Poem Reader and the Similarity Explorer.

Algorithm Basis Description
TF-IDF Lemma Lemma-level Cosine similarity on TF-IDF vectors of lemmatized poem texts. Captures thematic similarity through shared vocabulary, weighted by corpus-level term importance.
Wordform Overlap (Jaccard) Exact wordforms Jaccard index over raw wordform sets. Identifies poems sharing exact surface forms, useful for detecting formulaic lines and direct textual parallels.
Thematic (Translation) Cross-lingual Pivoted cosine similarity over English translations derived from DeepSeek annotations. Enables cross-lingual comparison between Estonian and Finnish poems via a shared semantic space.
Alignment Character n-gram Verse sequence alignment using character bigram cosine similarity and dynamic programming, from the FILTER project (Janicki, Kallio & Sarv 2023). Captures structural similarity — poems that follow the same verse order score high.
Semantic (GloVe) Embeddings GloVe 300-dimensional word vectors combined using SIF weighting (Arora et al. 2017) on English translations. Captures broad thematic resemblance even when poems use entirely different words. Cross-lingual by nature.
Verse Match Verse-level fusion Runs five verse-level algorithms (Jaccard, TF-IDF, Translation, CharBigram, Sentence) on individual verse pairs, aggregates best matches per verse via RRF, then combines into a single poem-level ranking.
Combined (RRF) Meta-ranking Reciprocal Rank Fusion across TF-IDF, Jaccard, Translation, and Alignment. Each algorithm’s ranking feeds one fused score; agreement pushes a poem up, but single-method matches can still appear. Scores shown as percentage of theoretical maximum.

Verse-level similarity algorithms

Five algorithms compare individual verse lines across the corpus. Each captures a different kind of textual similarity. Their results feed into tools like Verse Similarity, Verse Pairs, and the Geography maps, and are also aggregated into the poem-level Verse Match algorithm above.

Algorithm What it compares Description
Jaccard Shared words Counts how many exact words two verses share, divided by the total number of distinct words in both. High scores mean the verses use largely the same vocabulary.
TF-IDF Rare words Like Jaccard, but weights rare words more heavily than common ones. Two verses sharing an unusual word score higher than two sharing a word that appears everywhere.
Translation Cross-lingual meaning Compares verses through their English translations, so it works across languages. Two verses with the same meaning score high even if the original words differ entirely.
CharBigram Letter patterns Compares how words look letter-by-letter using character bigrams (two-letter pairs). Catches similarities between dialect spellings or closely related word forms that exact-word methods miss.
Sentence Semantic meaning Compares verse meanings using all-MiniLM-L6-v2 sentence embeddings on English translations. Captures broad thematic resemblance even when the original words differ entirely. Cross-lingual by nature.

Geography metrics

The Geography section of the Similarity Explorer uses several metrics to measure how strongly two collection places are connected through shared verse material. The first four metrics rank the connections drawn in the Verse Network view; the Place Focus view’s two scores (Tradition Strength, Surprising Connections) are computed differently and described in the last two rows.

Metric Description
Normalized surprise How unexpected a verse connection is between two places: identical shared verse lines divided by the geometric mean of the two places’ distinct verse-line counts. A high score means two places share more than their sizes alone would predict.
Combined count The total of shared verse connections between two places — identical lines plus retained near-match pairs, two different units, so treat it as a rough volume signal. Larger collections naturally produce higher counts.
Identical formulas only Counts only verse lines that appear word-for-word the same in both places.
Near-matches only Counts verse pairs that are algorithmically similar but not identical — dialect variants, minor word substitutions, and partially shared lines.
Tradition Strength How much poetic material two places share, used in the Place Focus view: every unique verse pair the places share contributes its best similarity score across four verse-level algorithms, and the scores are summed — match strength times volume, not a plain count.
Surprising Connections Used in the Place Focus view: the same summed score divided by the geometric mean of the two places’ total collected verse lines, so a small place that shares a lot ranks high.

← Back to About

About RunoVerse · Back to Lexicon