sinovation / zen2 Goto Github PK

Dear authors,

I have seen the ZEN2 achieves great success on the NLPCC-DBQA task from ERNIE 3 paper, I have gone through GitHub but found no answer yet for the DBQA task except the using NLPCC-DBQA dataset to do sequence-level classification fine-tune. I wonder can you give me some hint on how you apply ZEN2 to the DBQA task i.e. semantic parsing or information retrieving (embedding)? Thanks a lot.

Best Regards,
Zhang Yong

Excuse me, when I use the fine-tuned msra_ner model to fine-tune other dateset(like weibo, resume..), it return error sometimes, like
`Traceback (most recent call last):

File "run_token_level_ft_task.py", line 1263, in
--main()
File "run_token_level_ft_task.py", line 1240, in main
--result = evaluate(args, model, tokenizer, ngram_dict, processor, label_list, device, mode="dev")
File "run_token_level_ft_task.py", line 976, in evaluate
--temp_2.append(label_map[logits[i][j]])
KeyError: 0
`
the label_map doesn't have the key-"0", but I get it, (I dont modify the code) and my paratemers is

os.system("python run_token_level_ft_task.py --task_name weibo --seed 123 " "--do_eval --eval_test --do_lower_case --data_dir ./dataset/weibo " "--init_checkpoint dataset/zh_zen_base_2.0_ft_msraner --max_seq_length 128 --output_dir ./result " "--learning_rate 1e-5 " "--train_batch_size 8 --num_train_epochs 30 --warmup_proportion 0.1")

If possible, could you check this question, thanks!!