LearnToCompareText

implement "StructuredSelfAttention" + "RelationNetwork" for few shot learning of text

step

python Util.py python fewshot_main.py

how

more data via manual annotation or data augmentation
more features via transfer learning
more train via meta learning
less parameters and other robust

performance

30seq 10000step 300dim minum100shot
embeeding +cosine  0.54   
embedding+ [bi]GRU + cosine 0.59/0.61
embedding+ [bi]LSTM + cosine 0.63/0.61
embedding+ attn BiGRU + cosine 0.77
embedding+ attn BiLSTM + cosine 0.76
embedding+ attn BiGRU + cosine + data arguementation 0.79
embedding+ attn BiLSTM + concat not converge
bert... tokenize in task_generator

data augmentation √
seqquence length √
pretrain √
less parameters √
c-way-k-shot √

Reference：

1. Few-Shot Text Classification with Induction Network https://arxiv.org/abs/1902.10482
2. Learning to Compare: Relation Network for Few-Shot Learning https://arxiv.org/abs/1711.06025 https://github.com/floodsung/LearningToCompare_FSL
3. A Structured Self-attentive Sentence Embedding https://arxiv.org/abs/1703.03130 https://github.com/kaushalshetty/Structured-Self-Attention
4. corpus https://github.com/fate233/toutiao-multilevel-text-classfication-dataset
5. char_vector https://github.com/Embedding/Chinese-Word-Vectors