This is the source code for open_domain_entity_state_tracking

package version

NVIdIa Quadro rTX 8000 48G

python 3.6.5

torch version: 1.4.0+cu100

transformers :4.1.1

1-Data Progress

1. to correct the data in OpenPI (Data_progress/0_Data_Correction)
2. how to retrival the knowledge in the ConcpetNet in hop-N (Data_progress/1_get_kowledge_from_KG)
3. to create two selections (entity selection, attribute selection) (Data_progress/2_selection_model)
4. to build the entity-attribute knowledge graph by selected entities/attribute knowledge (Data_progress/3_combine_entity_attr_to_graph)
5. others (how to get the entity/attribute embedding) in step 4). (Data_progress4_get_E_embedding) Note: above file is in the Google Driver Data progress

2-Entity and Attr Selection Building

In this section, a pre-trained model with the triple loss function is used to create the entity and attr selection. the source code is in selection model. please uses the default parameters in the scripts. How to run? In entity or attr selection model, please run:

CUDA_VISIBLE_DEVICES=0 python train.py

3-KIEST w/o ESC

This is the variants of KIEST without the reward. Full model, please check KIEST w/o ESC

3.1-Train: please enter the 2_KIEST_constraint and run:

CUDA_VISIBLE_DEVICES=2 python training/run_trainer.py --output_dir=training_output --model_type=t5-large --continue_from_dir=/continue_from_dir --model_name_or_path=t5-large --do_train --train_data_file=data/formatted_for_gpt2/train.jsonl --per_gpu_train_batch_size 8 --per_gpu_eval_batch_size 16 --overwrite_output_dir --length 100 --block_size 300 --save_total_limit 3 --save_steps 5000 --learning_rate 0.00005 --overridden_model_configs '{"resid_pdrop": 0.1, "attn_dropout": 0.1}' --weight_decay 0.0 --num_train_epochs 50 --seed 42

Note: the trained checkpoinks file is in training_output.

3.2-Generation: Please get into the 2_KIEST_constraint file and run,

bash scripts/predictions_bash.sh

you can change your own paramters in the predictions_bash.sh. The generated file 2_KIEST_constraint\data\prediction_format.jsonl is same as the file test_our_model\prediction_format_wo_ESC.jsonl

3.3-Evaluation： please get into 6_test_our_model, and run:

python 1_change_file.py
python 2_simple_eval.py

3.4-Results：

Note: the BLEU is the BLEU-2, ROUGE is the ROUGE-L.

4-Classification Reward:

get into the 3_reward_clasification and run:

CUDA_VISIBLE_DEVICES=0 python train.py

5-KIEST:

Full files please check 4_KIEST_cosntrain_reward

5.1-Train: please enter the 4_KIEST_cosntrain_reward and run:

CUDA_VISIBLE_DEVICES=1 python training/run_trainer.py --output_dir=training_output --model_type=t5-large --continue_from_dir=/continue_from_dir --model_name_or_path=t5-large --do_train --train_data_file=data/formatted_for_gpt2/train.jsonl --per_gpu_train_batch_size 8 --per_gpu_eval_batch_size 16 --overwrite_output_dir --length 100 --block_size 300 --save_total_limit 3 --save_steps 5000 --learning_rate 0.00005 --overridden_model_configs '{"resid_pdrop": 0.1, "attn_dropout": 0.1}' --weight_decay 0.0 --num_train_epochs 10 --seed 10

Note: the trained checkpoinks file is in training_output.

5.2-Generation: Please get into the 4_KIEST_cosntrain_reward file and run,

bash scripts/predictions_bash.sh

5.3-Evaluation： please get into 6_test_our_model, and run:

python 1_change_file.py
python 2_simple_eval.py (the input file is prediction_format_constraint_reward_define_0.1.jsonl)

5.4-Results：

Note: the BLEU is the BLEU-2, ROUGE is the ROUGE-L.