hskimim / leam-pytorch

pytorch implementation of leam

leam-pytorch

Pytorch implementation Code implementation of Joint Embedding of Words and Labels for Text Classification(LEAM)

Data description

• yelp 2015
  • download source
  • label : stars [0..5]
  • only used 1,000,000
  • label imbalanced is mitigated by over-sampling
    • each # of label is sampled to 289394

Model description

• please refer the src/model.py in detail

• performance

  • VAL ACC : 0.713 | VAL LOSS : 0.702 at epoch 5

• word embedding(V) distribution

  • V has high-dimension [batch_size, max_length, hidden_dim]
  • I reduced dimension by averaging [batch_size, hidden_dim] (document vector)
  • I applied t-sne to reduce the embedding dimension [batch_size, 2]

• label embedding(C) cosine similarity matrix
  • C has dimension [label_num(6 in yelp review), hidden_dim]
  • cosine similarity is caculated