libing125

After changing the code, I ran slot filling on dstc8 fulll data, got 86.93 accuracy, lower than 90.05(the 'BERT' setting) reported in your article. I also tried run few-shot setting...

my script: ``` export BERT_MODEL_DIR=pretrained_models/bert-base-uncased export BERT_VOCAB_PATH=$BERT_MODEL_DIR/vocab.txt CUDA_VISIBLE_DEVICES=3 python3 run.py \ --train_data_path data_utils/dialoglue/dstc8_sgd/train.json \ --val_data_path data_utils/dialoglue/dstc8_sgd/val.json \ --test_data_path data_utils/dialoglue/dstc8_sgd/test.json \ --token_vocab_path $BERT_MODEL_DIR/vocab.txt \ --train_batch_size 64 --dropout 0.1 --num_epochs 100 --learning_rate...

My scikit-learn and numpy version are not identical to yours. I updated them and got 44.61(45.05 reported) few-shot accuracy, But only got 85.8 on full dstc8 data. I‘m confused.

I think they are the same.

之前 data/multiwoz/ontology.json 有问题，修复了 修改了sumbt的接口