On semeval performance
libing125 opened this issue · comments
I run your code with
CUDA_VISIBLE_DEVICES=0,1,2,3 python3 run_classifier_TABSA.py --task_name semeval_QA_B --data_dir data/semeval2014/bert-pair --vocab_file uncased_L-12_H-768_A-12/vocab.txt --bert_config_file uncased_L-12_H-768_A-12/bert_config.json --init_checkpoint uncased_L-12_H-768_A-12/pytorch_model.bin --eval_test --do_lower_case --max_seq_length 512 --train_batch_size 24 --learning_rate 2e-5 --num_train_epochs 12.0 --output_dir results_v2/semeval_result/QA_B --seed 42
and get best performance at epoch 10
aspect_P = 0.9109792284866469
aspect_R = 0.8985365853658537
aspect_F = 0.9047151277013752
sentiment_Acc_4_classes = 0.8439024390243902
sentiment_Acc_3_classes = 0.8735868448098664
sentiment_Acc_2_classes = 0.9306029579067122
But it should be
BERT-pair-QA-B 85.9 89.9 95.6
as in your article.
What should I do to reproduce your results? Thanks.
The recommended epoch for fine-tuning BERT model is 3 or 4.
Besides, you can try to change the random seed.