Skip to content

Latest commit

 

History

History
56 lines (42 loc) · 3.74 KB

README.md

File metadata and controls

56 lines (42 loc) · 3.74 KB
Raw

Introduction This project targets problems of insufficient data in text classification tasks. By using some few-shot learning tricks (ProtoNet, etc.), performance on tasks sees improvement, and has potential to furthur improve, but the convergence speed for ProtoNet+bert is much slower than normal Bert finetuning, and GPU memory is also a key limitation on its improvement (cannot set large number of supports at evaluation time, #TODO to fix this in the future)

*文本小样本多分类模型(仅测试过短文本), 目前用bert初始化, 可换用sentence-bert做初始化, 效果更佳

Classification Models

  1. ProtoNet+Bert (optimized for fewshot, can achieve better performance on some small dataset)
  2. Ordinary Bert classification (for normal dataset, also works for fewshot thanks to the strength of BERT pretraining)
  3. A Mysterious Algorithm from my colleague (optimized for matching tasks, do not train this for normal classification tasks, just for experimental purporse, just for fun)

Usage:

  1. put your data into ./data folder
  2. write your own script (or use some pre-given function in data_formatter.py) to format your training/evaluation data into "sentence and its label separated by tab" per line
  3. modify configuration in conf/config.py under the Config class for your chosen model,
    • Mandatory settings:
      • for Bert classifier: set number of classes and max sentence length,
      • for ProtoNet: set "k" and "shot", k must be between 20% to 100% of total number of classes, shot commonly between 2 and 10 depending on datasize
    • Optional settings:
      • for Bert classifier: batch_size
      • for ProtoNet: n_support, eval_n_support (number of supporting samples for each class, read the paper on ProtoNet for more details), you can just leave them unchanged, the bigger the better, but may exceeds GPU memory limits, especially at evaluation time, when number of classes is big.
      • general settings: learning rate, warmup, paths to essential data/modelfiles, device, etc..
  4. Alternatively, if you are sick of modifying the config file, or you want to train multiple models with different configs, you can just use <python scripts/api.py> directly, all kinds of settings can be re-defined here, overriding what's in config.py. type <python scripts/api.py -h> for more details.
  5. choose to run from three shell script on your demand
  6. predict with the other three shell scripts, don't forget to check all kinds of load paths before running

Requirements: pytorch, transformers, pytorch_pretrained_bert, keras, sklearn, etc..

Note: Recommended hyperparameters are left as they are in conf/config.py except those that are task specific. All experiments are using bert-chinese-base, not tested for other languages, but you can always try it (remember to change bert_type in config).

TODO:

  1. support unlimited number of supports at evaluation/prediction time
  2. support Meta-Learning
  3. replace Euclidean distance with RE2 and BCEloss

-------------------------------------------------------------------------------------------------------:

ProtoNet+Bert Bert Training size Test size Balanced Class Count
Intent Classification (downsampled to 1%) 88.3% 87.4% ≈60*15 1333 True 15
Intent Classification (downsampled to 10%) 91.9% 91.7% ≈600*15 1333 True 15
Intent Classification >93.7%(too slow to train) 94.6%(?) ≈6000*15 1333 True 15
Anonymous Dataset 1 87.8% 87.2% 3200 352 False 86
Anonymous Dataset 2 84.9% 84.3% 1300 434 False 20
Anonymous Dataset 3 88.1% 83.9% 5000 320 True 68