What is the Semantic Explorer?

This tool lets you explore 696,310 annotated wordforms across 51 semantic categories (32 entity + 19 emotion) in the Finnic runosong corpus. Annotations are drawn from three merged systems:

• ML Pipeline (P) — 13-method machine learning pipeline using NER (GLiNER, WordNet, dictionary, morphological), emotion detection (GoEmotions, SetFit, NRC EmoLex, DepecheMood), and lemma propagation. Covers 544K wordforms.
• Thesaurus (T) — 25-domain keyword matching on English translations. Adds 401K wordforms with high precision for verbs and abstract concepts.
• EKKD Emotion Dictionary (E) — Curated Estonian runosong emotion vocabulary (1,937 lemmas, 4 classification tiers: E=core, M=derivative, A=artifact, J=neighbor). Adds 29K emotion-specific wordforms.

138K wordforms are cross-validated (confirmed by multiple systems), increasing confidence.

Tabs

• Overview — Corpus-wide stats, provenance breakdown (which system contributed what), spotlight cards, category bar chart, method donut, co-occurrences.
• Browse — Select categories from the sidebar, explore wordforms grouped by lemma. Use filters for language (ET/FI), confidence (High/Med/Low), source (Pipeline/Thesaurus/EKKD), and method. Click seed keyword pills to filter by defining concepts. Use the Compare button for side-by-side category analysis.
• Geography — Map showing where semantic categories concentrate geographically across collection places.
• Cross-lingual — Butterfly chart comparing Estonian vs Finnish distribution per category. Click to expand top wordforms.
• Discoveries — Newly discovered emotion vocabulary (wordforms not previously known from verse substitution analysis), major category table, method overlap matrix, geographic outliers.

Browse Features

• Keyword search — Type an English keyword (e.g., "mother", "horse") in the green search bar to find matching categories.
• Seed pills — Clickable keyword chips under the category header. Click to filter the table to wordforms matching that seed.
• Source badges — P = ML Pipeline, T = Thesaurus, E = EKKD. Shown next to each lemma's methods.
• Compare — Click the orange "Compare..." button to select a second category and see shared lemmas, seeds, and stats side by side.
• Export CSV — Download the current filtered view as a tab-separated file.

Confidence Levels

• High (≥0.70) — Multiple methods agree or expert-curated (EKKD E-class).
• Medium (0.40–0.69) — At least one reliable method with moderate support.
• Low (<0.40) — Single weak signal; includes EKKD J-class (neighbor words) at 0.39.