📄 WordNet Affect: an Affective Extension of WordNet

Summary

Earlier emotion lexicons mostly worked with words rather than senses. They often relied on factor analysis or multidimensional scaling, which cannot tell surprise the feeling apart from surprise the event. The authors address this by building WordNet-Affect, a sense-level affective lexicon layered on WordNet without changing WordNet's structure. Following the design of WordNet Domains, chosen synsets get one or more affective labels (a-labels), such as EMOTION, MOOD, TRAIT or EMOTIONAL RESPONSE.

Construction had two phases. First, a hand-built database called AFFECT, holding terms with lexical and psychological attributes, was mapped onto WordNet senses. Non-affective synsets were dropped, synonyms were automatically checked for conflicting affective values, and labels were projected and then completed by hand. Second, this core was expanded through WordNet relations judged to keep affective meaning.

The result covers 2,874 synsets and 4,787 words and is free for research use. It gives emotional agents, empathetic text generation, affect detection and computational humour a sense-aware inventory of affective vocabulary.

Key points

• The seed database AFFECT contains 1,903 terms: 539 nouns, 517 adjectives, 238 verbs and 15 adverbs. Each term has English–Italian links, definitions, synonyms and antonyms, a POSR cross-POS link (e.g. cheerful → cheerfulness, cheer up, cheerfully) and an ORTONY category slot.
• WordNet-Affect totals 2,874 synsets and 4,787 words:
  • 763 noun synsets
  • 1,462 adjective synsets
  • 322 verb synsets
  • 327 adverb synsets
• Expanding through relations added 1,297 synsets. The largest contributors were similar-to (573), derived-from (308) and antonym (169).
• Some relations were treated as "reliable", meaning they were assumed to preserve affect and were added automatically: antonymy, similarity, derived-from, pertains-to, attribute and also-see. Others were treated as only partly reliable and needed manual filtering: hyperonymy, entailment, causes and verb-group.
• The available affective labels (a-labels) are:
  • MOOD, TRAIT, EMOTION
  • EDONIC SIGNAL, PHYSICAL STATE, COGNITIVE STATE
  • EMOTIONAL RESPONSE, EMOTION-ELICITING SITUATION
  • ATTITUDE, SENSATION, BEHAVIOUR

Interesting ideas

• I1. Annotate, don't restructure. Adding a separate label hierarchy on top of existing synsets, as WordNet Domains does, avoids forking the base resource. This approach could support a Ukrainian wordnet with sentiment or domain layers that stays compatible with the original.
• I2. Tier relations by trust. Relations were split into ones assumed to preserve the target property, used for automatic expansion, and ones needing human review. This sensible way to spend annotation effort applies to any lexicon grown by projection.
• I3. Check consistency within synsets. Synonyms carrying incompatible affective values were flagged automatically, which both catches errors and tests whether synsets suit the property. This is an easy validation step for any word-level-to-sense-level label transfer.
• I4. Link categories across parts of speech. The POSR-style link groups adjective, noun, verb and adverb forms of one psychological category. This lets labels spread across POS and helps generation systems pick alternative forms of a concept.
• I5. Use valence opposition for humour. For ironic acronym re-analysis, expansion words are deliberately chosen with valence opposite to the original. This shows how a labelled lexicon can steer creative generation, not just classification.