What this shows

A network of similar verse lines from the Finnic runosong corpus (4.32M verses, 289,702 poems). The seed verse (green node) is at the center. Connected verses share similar wording, meaning, or translation.

How similarity is calculated

Five algorithms independently find similar verses:

• Jaccard — shared exact wordforms between two verse lines
• TF-IDF — shared lemmas (dictionary forms), weighted by rarity
• Translation — shared English translations, enabling cross-lingual Estonian↔Finnish matches
• CharBigram — character bigram overlap, capturing surface-level textual similarity
• Sentence — sentence embedding similarity using all-MiniLM-L6-v2 on English translations

These are combined via Reciprocal Rank Fusion (RRF): if a verse ranks high in multiple algorithms, its combined score is boosted. The percentage shown is the RRF score normalized to 0–100%.

Reading the graph

• Node size = occurrence frequency (log-scaled: how many times this exact verse text appears across the corpus)
• Node color: green = seed verse, blue = Estonian, orange = Finnish
• Edge color: green = seed connection, purple = inter-neighbor link, gray = hop-2+
• Edge thickness = similarity strength

What are hops?

• Hop-1: verses directly similar to the seed (the seed's nearest neighbors)
• Hop-2: verses similar to hop-1 nodes (neighbors of neighbors)
• Higher hops extend further. Each hop level discovers more distant connections.

Controls

Control	What it does
Hop-1	How many direct neighbors to show (5–20)
Per-hop	How many new nodes each parent adds at hop-2+ (1–5)
Max hops	Network depth (1–4). Default 2 for performance.
Min score	Hide weak connections below this RRF threshold
Max nodes	Cap total visible nodes (15–200)
ET / FI	Filter by language (Estonian / Finnish)
+ Add hop	Extend network by one more hop level without full reload
Show outer hops	Toggle visibility of hop-2+ nodes
Show labels	Toggle verse text labels on/off
Zoom slider	Zoom in/out (20–500%)

Interactions

• Click a node → navigate to that verse's network (re-centers on it)
• Click the seed node (green) → show info panel with links to Similarity Explorer and Poem Reader
• Double-click a node → expand it in-place (load its neighbors without navigating away)
• Double-click empty space → reset zoom to default
• Drag a node to reposition it; scroll to zoom; drag background to pan
• Nodes with a + overlay can be expanded by double-click

Exporting data

Open "More controls" and click "⬇ Export CSV" to download all visible verses and edges with per-algorithm similarity scores and shared wordforms.

Colour modes

The Colour dropdown switches how nodes are coloured:

• Language (default) — blue = Estonian, orange = Finnish, green = seed.
• Poem formulaicity — a poem-level colour: each node is shaded on a pale-to-red ramp by its parent poem's formulaicity — the share of that whole poem's verses that recur as shared formulae across the corpus (FILTER cluster matching). Because the score belongs to the poem, not the line, every verse from the same poem gets the same colour: it tells you how formula-heavy the source poem is, not whether this one line is itself a stock phrase. Pale yellow = rare wording, deep red = highly formulaic stock phrasing, grey = no score. Coverage differs by tradition (ET ~43%, FI ~53%), so compare formulaicity within a tradition, not across.
• Verse formulaicity — a line-level colour showing how much each verse is a recurring stock phrase, computed live in your browser. The main ingredient is how widely the line recurs across the corpus: the system counts how many other poems contain a matching verse — where "matching" includes spelling, dialect, and small wording variants, not just identical copies — pooled across all five similarity measures (shared wordforms, weighted wordforms, translation, character n-grams, and sentence embedding). The more poems echo the line, the more formulaic it is; that count saturates once a line recurs in about 50 poems. Two smaller factors refine the score: a line ranks higher if its single closest parallel is a near-exact twin (rather than only loose echoes), and very short lines are damped because they match by chance. Because the score belongs to the line, two verses from the same poem can be coloured differently (unlike Poem formulaicity). The number in a node's tooltip (e.g. 91 / 100) is an absolute, corpus-wide value — 100 = near the most formulaic lines anywhere — and never changes between graphs. The colour, by contrast, is relative to this graph: the pale-to-deep ramp is stretched between the least- and most-formulaic verses currently on screen and re-stretches as the network grows, so a shade means "formulaic compared with the rest of this network" and is not comparable between different seeds — for a comparable value, read the tooltip number. Pale magenta = low, deep magenta = high within this view, near-black = a top-capped line, grey = not scored (too short, no parallel, or data still loading).

Network statistics

The Stats button toggles a bar of live figures for the visible network: node count (with hop breakdown), the cross-lingual share of edges (Estonian↔Finnish links), the number of distinct places, the Estonian/Finnish split, the average formulaicity of the visible nodes (in either Poem or Verse formulaicity mode), and the mean RRF similarity of the edges.

Geographic map

The Map button opens a mini-map of where the visible verses were collected. Circle size = number of verses from that place; circle colour follows the current colour mode (language, poem formulaicity, or verse formulaicity); thin lines link places joined by a verse-similarity edge. Click a circle to highlight that place's verses on the graph (an orange ring); click it again to clear.

Keyboard shortcuts

Key	Action
/	Focus the verse search box
?	Toggle this help panel
S	Toggle the statistics panel
G	Toggle the geographic map
C	Copy a share link to this network
X	Export the visible network as CSV
R	Jump to a random verse in the network
Esc	Close the help panel, search results, or map