×

About the Emotion Vocabulary Browser

This page presents the emotion-related vocabulary of Finnic runosongs (regilaulud) — 63,871 word forms grouped into 38 families across 26 semantic domains. The dataset was assembled by combining 12 detection methods. Most operate on English translations of lemmatized runosong texts produced by the DeepSeek R1 LLM; the substitution test is the one method that works directly on the Estonian verse templates themselves.

How the vocabulary was built — 12 detection methods

• Lemmatization — EstNLTK + dialect corpus, Estonian and Finnish dictionaries, and LLM-based lemmatization (Pivovarova et al. 2025). Four complementary lemmatizers are combined to map each surface word form to a canonical lemma.
• Substitution test (Harris 1954) — for each seed word, the word is removed from every verse it appears in, producing a verse template. Every other word that fills the same template position anywhere in the corpus becomes a candidate. Template specificity is IDF-weighted. 26 emotion seed words were expanded to depth 5.
• LLM emotion extraction (Claude Opus) — extracts which words in a verse context actually denote emotion, filtering candidates from the substitution test into the E / M / A / J / N categories below.
• SetFit (Tunstall et al. 2022) — a text classifier fine-tuned on ~20,000 emotion words previously found via the substitution test + LLM. Predicts emotion domains for all 1.2M translated word forms.
• GoEmotions (Demszky et al. 2020) — a RoBERTa classifier that scores the translation of each runosong word for 28 emotion categories.
• NRC EmoLex — Plutchik-style emotion scores for 14,000 English words, plus native Estonian and Finnish NRC lexicons matched directly on the source forms.
• DepecheMood++ — crowd-sourced emotion scores for 100,000+ English lemmas, transferred onto runosong words through their translations.
• GLiNER (Zaratiana et al. 2024) — zero-shot named-entity recognition used to filter out concrete entities (persons, animals, objects) so they do not receive spurious emotion labels.
• WordNet (Princeton / NLTK) — semantic categorization of translations into entity types, blocking emotion annotations on words classified as animals, plants, foods, etc.
• Lemma-based expansion — once a lemma is confirmed emotional, all of its paradigmatic word forms inherit the label. Modern Estonian emotion dictionaries (EKKD emotsioonisõnastik, v3–v7) are cross-referenced to validate expansions.

All methods are merged with consensus scoring: a word form is confirmed as emotional only when ≥2 independent methods agree. Each family is then reviewed by Claude Opus against actual verse evidence, word-form translations, and corpus frequencies, and three layers of corrections are applied (per-form moves, lemma-level merges, recovered forms). Final corrections propagate to the RunoVerse lexicon and poem index.

Classification categories

E = core emotion word (e.g., rõõm ‘joy’). M = derivation of an emotion word (e.g., rõõmustama ‘to rejoice’). A = dialect variant or lemmatization error merged to its correct family (e.g., reem → rõõm). J = contextually related but not an emotion itself (e.g., laul ‘song’, süda ‘heart’). N (noise) candidates are excluded from the browser.

Two different “substitution tests” on RunoVerse

This page uses the emotion-pipeline substitution test: a seed word is removed from its verse to form a template, and every word in the corpus that fills the same template position becomes a candidate for the seed’s emotion family. Templates are scored by IDF so that specific, low-frequency slots count more than generic ones. The method is seed-driven and iterates to depth 5.

The separate Substitution Explorer is built from the main RunoVerse similarity pipeline: 701K substitution pairs produced by cross-poem (XP) and within-poem parallelism (CP) algorithms. It is not seed-driven and not restricted to emotion words — it surfaces any two words that substitute for each other in parallel verse lines. The two datasets are complementary (only ~13.6% overlap).

Using this browser

• Sidebar: Browse emotion families grouped by semantic domain. Click a family to view its members.
• Search: Search by lemma, word form, or English translation. Matches filter the sidebar in real time.
• Word forms: Blue forms (~45K of 64K) were directly confirmed by the substitution test; gray forms entered through translation-based methods or lemma-based expansion. Frequency badges show corpus count.
• Verse examples: Hover over a word form to see an example verse. Click to open the word’s detailed view with poem contexts.
• Sort: Word forms can be sorted alphabetically (A-Z), by corpus frequency (Freq), or by template IDF weight (IDF — how specific the substitution templates are).
• Export: Use the CSV or JSON export buttons on the landing page.

Translation cognates

Each family page includes translation cognates — Estonian lemmas from the runosong corpus that share English translations with the family’s members. The RunoVerse lexicon (242K glosses, 1.5M mappings from DeepSeek AI translations) is inverted to build per-lemma gloss profiles. For each family, lemmas whose translation profiles overlap with the family’s are scored using an IDF-weighted overlap coefficient: common glosses (‘to’, ‘little’) are downweighted via idf = log(1 + N/(1 + df)), while specific glosses (‘contempt’, ‘mock’) count more. Candidates require ≥2 shared glosses and a minimum IDF sum. This discovers semantically related words that the substitution test may have missed.

Source data

Corpus: Estonian (ERAB) and Finnish (SKVR, JR) runosong collections (7.3M + 7.4M tokens, 451K + 701K unique forms). Translations from the RunoVerse lexicon (DeepSeek R1 AI translations, 1.36M entries). BERT embeddings fine-tuned on runosong texts (190K words × 768 dimensions) are used for cognate scoring. Full pipeline documentation: EMOTION_LEXICON_PIPELINE_DOCUMENTATION.md.

Note: This vocabulary is computationally derived and AI-reviewed. It is intended as a research tool, not a definitive classification. Dialectal and archaic Estonian and Finnish (13th–19th century) presents inherent lemmatization challenges; the estimated false-positive rate after consensus + Claude review is ~9%.