BERT微调

Model Architecture

• Predict intent and slot at the same time from one BERT model (=Joint model)

• total_loss = intent_loss + coef * slot_loss (Change coef with --slot_loss_coef option)

• If you want to use CRF layer, give --use_crf option

Dependencies

• python>=3.6

• torch==1.6.0

• transformers==3.0.2

• seqeval==0.0.12

• pytorch-crf==0.7.2

Dataset

	Train	Dev	Test	Intent Labels	Slot Labels
ATIS	4,478	500	893	21	120
Snips	13,084	700	700	7	72

• The number of labels are based on the train dataset.

• Add UNK for labels (For intent and slot labels which are only shown in dev and test dataset)

• Add PAD for slot label

Training & Evaluation

$ python3 main.py --task {task_name} \
                  --model_type {model_type} \
                  --model_dir {model_dir_name} \
                  --do_train --do_eval \
                  --use_crf

# For ATIS
$ python3 main.py --task atis \
                  --model_type bert \
                  --model_dir atis_model \
                  --do_train --do_eval
# For Snips
$ python3 main.py --task snips \
                  --model_type bert \
                  --model_dir snips_model \
                  --do_train --do_eval

# python JointBERT/main.py --task atis --model_type bert --model_dir experiments/jointbert_0 --do_train --do_eval --train_batch_size 2

Prediction

$ python3 predict.py --input_file {INPUT_FILE_PATH} --output_file {OUTPUT_FILE_PATH} --model_dir {SAVED_CKPT_PATH}

Results

• Run 5 ~ 10 epochs (Record the best result)

• Only test with uncased model

• ALBERT xxlarge sometimes can't converge well for slot prediction.

		Intent acc (%)	Slot F1 (%)	Sentence acc (%)
Snips	BERT	99.14	96.90	93.00
	BERT + CRF	98.57	97.24	93.57
	ALBERT + CRF	99.00	96.55	92.57
ATIS	BERT	97.87	95.59	88.24
	BERT + CRF	97.98	95.93	88.58