What is the Semantic Explorer?

This tool lets you explore 339K+ annotated wordforms in the Finnic runosong corpus through three semantic tiers, organized by reliability:

1. Thesaurus Domains (25) — Recommended

These 25 themes are the most reliable layer, because a word is sorted into a theme through its dictionary translation rather than by a machine-learning model trained on modern text.

How a word joins a theme. Each theme is defined by a hand-picked list of English keywords, called its seeds — for example, Family & Kinship is defined by mother, father, sister, bride, orphan, widow, kin… Every word in the corpus has already been given an English meaning in the dictionary layer, and a word is placed in a theme whenever that meaning matches one of the theme's seed keywords. So the Estonian ema and the Finnish emo both land in Family & Kinship because each translates as “mother.” A word can belong to several themes at once if its meaning fits more than one.

The link runs by matching the English gloss against a hand-picked list of theme words, so it reaches archaic and dialectal vocabulary that has a gloss. Accuracy depends on that gloss: a word with several senses can land in the wrong theme. The trade-off is breadth over nuance: themes are wide thematic buckets (Animals, Water & Sea, War & Violence…), not fine-grained word senses.

• Each theme lists its top words with frequency, language (ET/FI/both), and part of speech
• Click a seed-keyword pill to narrow a theme to a single defining concept
• Built from thesaurus_index.json — 25 themes, 40K+ keyword links

2. Emotion Categories (26) — Reliable

These emotion categories come from the Runosong Emotion Lexicon — a vocabulary of feeling-words discovered inside the songs themselves rather than predicted by a model trained on modern language.

How an emotion word is found. The core method is a substitution test. Runosong reuses verse material heavily, so the same verse recurs with a single word swapped — rõõmu suurem ümberikku stands beside leina suurem ümberikku and muret suurem ümberikku. Words that keep filling the same slot in such a verse template are taken to be semantically related. The search begins from 26 hand-chosen seed words (joy, grief, anger, fear, love…) and gathers the words that share their slots; an AI reviewer then sorts each candidate into a core emotion word, a derived form, a misattributed form (an artifact or dialectal spelling that really belongs to another word), or a merely neighbouring word, and the confirmed emotion words become seeds for the next round. The surviving vocabulary is grouped into the categories shown here.

Because the evidence comes from inside the corpus, this layer holds up on the archaic, dialectal vocabulary where modern-text models stumble. The categories are then cross-checked against several modern sentiment classifiers, which raises confidence wherever independent methods agree.

• Emotion lexicon (E) — the Runosong Emotion Lexicon above, with each word marked as a core emotion word, a derivation, a misattributed form, or a neighbour
• ML Pipeline (P) — corroborating sentiment classifiers (GoEmotions, SetFit, NRC EmoLex, DepecheMood) and lemma propagation
• Data consistent with all other emotion pages (Emotion Families, Emotional Geography, Emotion Canvas)

3. NLP Entity Categories (26) — Experimental

Entity categories from NLP models (GLiNER NER, WordNet, thesaurus keyword matching, morphological detection, lemma propagation). These models were trained on modern English/general text and may have lower accuracy on archaic dialectal runosong vocabulary. Results are preliminary — use with caution.

• ML Pipeline (P) — Multi-method NER + WordNet + dictionary + morphological detection
• Thesaurus (T) — 25-domain keyword matching on English translations
• Cross-validation between sources increases confidence

Tabs

• Overview — Corpus-wide stats, provenance breakdown, category bar chart, method donut, co-occurrences.
• Browse — Three-tier sidebar: thesaurus domains (green), emotions (purple), NLP entities (grey/collapsed). Click to explore wordforms, filter by language/confidence/source/method.
• Geography — Map showing where semantic categories concentrate geographically.
• Cross-lingual — Butterfly chart comparing Estonian vs Finnish per category.
• Discoveries — Newly discovered emotion vocabulary, method overlap matrix, geographic outliers.

Confidence Levels

• High (≥0.70) — Several methods agree, or the lexicon marks the word as a core emotion word, a derivation, or a misattributed form (classes E/M/A, scored 0.92 / 0.85 / 0.75).
• Medium (0.40–0.69) — At least one reliable method with moderate support.
• Low (<0.40) — A single weak signal, such as a neighbouring word (class J) at 0.39.