Reorganization of Adjectives

[jmccrae]

Currently adjectives in OEWN are organized in the 'dumbbell' model, which creates some issues, most notably #35, but also issues due to having 'satellite adjective' as a different type of part-of-speech.

This proposal would eliminate the dumbbell model and replace it with an organization of adjectives that aims to connect the adjective hierarchy with the noun and verb hierarchy and contribute to the goal of #172.

The first part of this proposal is to introduce new relations based on the intuition that many adjectives are morphologically related to a verb or noun. These relations will be at a synset-level and so will be semantic even though they are closely related to syntactic derivations.

• present: relates an adjective to a verb and means "performing the verb currently or repeatedly", this is closely related to the present participle
  • Examples: demoralising, inductive, revealing
• resultant: relates an adjective to a verb and means "in a state after the action of the verb is completed", this is based on the past participle
  • Examples: categorized, cured
• potential: relates an adjective to a verb and means "capable of being the object of the verb". This corresponds to '-able' suffixes
  • Examples: deniable, ignescent (ignitable)
• lacking: relates an adjective to a noun and means "lacking in the noun". This corresponds to the '-less' suffix
  • Examples: hatless, sleeveless
• full_of: relates an adjective to a noun and means "full of the noun". This corresponds to the '-ful' suffix
  • Examples: faithful
• resembling: relates an adjective to a noun and means "resembling the noun". This corresponds to the '-like' suffix
  • Examples: fernlike, machinelike
• quality: relates to a noun, where the noun means "having the quality of the adjective". This corresponds to '-ness' or '-ity' suffixes on the noun
  • Examples: salty (to saltiness), efflorescent (to efflorescence)

An initial sample of 100 adjectives suggests these derivations cover 42% of adjective (present 9%, resultant 8%, potential 3%, lacking 2%, full_of <1%, resembling 4%, quality 16%)

For the remaining adjectives we can use relations that mostly already exist

• hypernym: Hypernyms don't exist for all adjectives but in many cases we can apply the pattern that if "if x is {hyponym} then x must necessarily be {hypernym}".
  • Examples: lighter-than-air → light, long-wooled → woolly
• antonym: Antonyms already exist but there are many missing, especially for the 'un-', 'in'- and 'mis-' prefixes
• pertainym: These already exist but are used in a limited manner. We can expand this with more general cases, e.g., 'sisterly' pertains to 'sister'
• scalar_quality: This would be a new relation indicating the adjective has some value on the scale of a noun, e.g., 'hot' to 'temperature'. This has already been proposed in other wordnets

Overall the most important of these from the sample appears to be hypernyms (35%), followed by antonyms (11%), pertainyms (7%) and scalars (4%)

I would also note that many synsets may have multiple of these relations

[arademaker]

These relations will be at a synset-level and so will be semantic even though they are closely related to syntactic derivations.

Why not making them sense-sense relations? Any special reason for keep them at synset level? The Morphosemantic Database from https://wordnet.princeton.edu/download/standoff-files are sense-sense relations.

[1313ou]

Are you referring to adjectives ? (because morphosemantic standoff files are indeed sense-to-sense relations but do not involve adjectives, afaik)

[jmccrae]

I think we want to use synset-level relations to express that these relations are purely semantic, that is that they exist even when there is not a morphological relation. This also means that many of these links do not correspond to morphological processes, e.g., 'visible' means can be seen, but is not morphologically related to the verb 'see'.