Part of codes and data for the paper Multi-turn Dialogue Reading Comprehension with Pivot Turns and Knowledge
KKT PART
(Our experiment environment for reference)
Python 3.7
PyTorch (1.0.0)
NLTK (3.4.5)
Raw data can be downloaded from the following urls:
DREAM: https://github.com/nlpdata/dream
ConceptNet: https://s3.amazonaws.com/conceptnet/downloads/2018/edges/conceptnet-assertions-5.6.0.csv.gz
We provides a processed version of DREAM and ConceptNet used in our experiment.
DREAM: in .KKT_code/dream_data/
We have written the relevance score and rank of each turn corresponding to QA in the json file.
We also have written the related knowledge IDs in ConceptNet for each example in the json file.
ConceptNet: in .KKT_code/conceptnet/
We have made a smaller version of ConceptNet, which only contains the English part.
Here is a simple example of our experiment:
- Get knowledge representation
python get_knowledge_representation.py
--data_dir=../conceptnet/
--model_type=albert
--model_name_or_path=albert-xxlarge-v2
--task_name=knowledge
--max_seq_length=40
--do_lower_case
--per_gpu_eval_batch_size=256 2. Run the model
python run_multiple_choice.py
--data_dir=../dream_data
--model_type=albert_all
--model_mode=kp
--model_name_or_path=albert-xxlarge-v2
--knowledge_file=../conceptnet/albert-xxlarge-v2_concept_knowledge_representations
--task_name=dream
--output_dir=../results/albert_xxlarge_dream
--warmup_steps=0
--max_seq_length=512
--top_k=6
--do_train
--do_eval
--do_test
--do_lower_case
--per_gpu_train_batch_size=2
--per_gpu_eval_batch_size=16
--learning_rate=1e-5
--num_train_epochs=2
--save_steps=383
--max_contexts_knowledge=30
--eval_all_checkpoints
--overwrite_output_dir
--overwrite_cacheThe results will be saved in ./results/albert_xxlarge_dream/my_eval_results.txt
Please kindly cite this paper if you use this repo.
@ARTICLE{9352490,
author={Zhang, Zhuosheng and Li, Junlong and Zhao, Hai},
journal={IEEE/ACM Transactions on Audio, Speech, and Language Processing},
title={Multi-Turn Dialogue Reading Comprehension With Pivot Turns and Knowledge},
year={2021},
volume={29},
number={},
pages={1161-1173},
doi={10.1109/TASLP.2021.3058616}
}