UWA (Unambiguous Word Annotations) - Dataset for WordNet Sense Embeddings (original) (raw)

PaperDon't Neglect the Obvious: On the Role of Unambiguous Words in Word Sense Disambiguation PDF

Datasets

• XML Format (w/SemCor)

  • SemCor+UWA 10 (recommended) (103 MB)
  • SemCor+UWA 1 (19 MB)

• JSON Format

  • UWA 10 (78 MB)
  • UWA 100 (514 MB)

• Extras

  • SemCor+UWA 10 Sense Counts (.tsv)

Embeddings

• RoBERTa (large)

  • SemCor+UWA10 Prop. (all sensekeys) (666 MB)
  • SemCor+UWA10 (123,087 sensekeys) (498 MB)
  • SemCor+UWA10 Prop. (all synsets) (391 MB)
  • WSD (1NN) Output Files

• BERT (large-cased)

  • SemCor+UWA10 Prop. (all sensekeys) (648 MB)
  • SemCor+UWA10 (123,087 sensekeys) (484 MB)
  • WSD (1NN) Output Files

Visualization of the Embedding Space

T-SNE comparison for synset embeddings that belong to the 'noun.food' supersense. See here (23MB) for visualization of all embeddings, or below for other WN groups. Using embeddings for synsets instead of sensekeys for clearer visualization. Synset embeddings learned by converting sensekey annotations in corresponding corpora.

Visualizations for each WordNet supersense

• adj.all
  14,435 synsets (58% vs. 67% inferred)
• adj.pert
  3,661 synsets (36% vs. 84% inferred)
• adj.ppl
  60 synsets (12% vs. 58% inferred)
• adv.all
  3,621 synsets (21% vs. 61% inferred)
• noun.act
  6,650 synsets (29% vs. 74% inferred)
• noun.animal
  7,509 synsets (21% vs. 96% inferred)
• noun.artifact
  11,587 synsets (23% vs. 81% inferred)
• noun.attribute
  3,039 synsets (29% vs. 70% inferred)
• noun.body
  2,016 synsets (16% vs. 85% inferred)
• noun.cognition
  2,964 synsets (23% vs. 71% inferred)
• noun.communication
  5,607 synsets (27% vs. 77% inferred)
• noun.event
  1,074 synsets (32% vs. 69% inferred)
• noun.feeling
  428 synsets (26% vs. 55% inferred)
• noun.food
  2,573 synsets (32% vs. 91% inferred)
• noun.group
  2,624 synsets (21% vs. 74% inferred)
• noun.location
  3,209 synsets (18% vs. 81% inferred)
• noun.motive
  42 synsets (12% vs. 69% inferred)
• noun.object
  1,545 synsets (18% vs. 83% inferred)
• noun.person
  11,087 synsets (18% vs. 85% inferred)
• noun.phenomenon
  641 synsets (16% vs. 71% inferred)
• noun.possession
  1,061 synsets (26% vs. 77% inferred)
• noun.process
  770 synsets (31% vs. 82% inferred)
• noun.quantity
  1,275 synsets (35% vs. 83% inferred)
• noun.relation
  437 synsets (29% vs. 75% inferred)
• noun.shape
  341 synsets (30% vs. 74% inferred)
• noun.state
  3,544 synsets (25% vs. 81% inferred)
• noun.substance
  2,983 synsets (14% vs. 85% inferred)
• noun.time
  1,028 synsets (16% vs. 69% inferred)
• noun.Tops
  51 synsets (18% vs. 35% inferred)
• verb.body
  547 synsets (39% vs. 61% inferred)
• verb.change
  2,383 synsets (48% vs. 64% inferred)
• verb.cognition
  695 synsets (33% vs. 46% inferred)
• verb.communication
  1,548 synsets (31% vs. 48% inferred)
• verb.competition
  459 synsets (42% vs. 62% inferred)
• verb.consumption
  243 synsets (35% vs. 54% inferred)
• verb.contact
  2,196 synsets (45% vs. 63% inferred)
• verb.creation
  694 synsets (40% vs. 61% inferred)
• verb.emotion
  343 synsets (33% vs. 46% inferred)
• verb.motion
  1,408 synsets (35% vs. 52% inferred)
• verb.perception
  461 synsets (38% vs. 52% inferred)
• verb.possession
  847 synsets (40% vs. 59% inferred)
• verb.social
  1,106 synsets (33% vs. 46% inferred)
• verb.stative
  756 synsets (36% vs. 43% inferred)
• verb.weather
  81 synsets (52% vs. 62% inferred)