domainKG

This is code for Experiment using domain specific KG.

We use the pytorch implementation for the EMNLP 2020 (Findings) paper: Connecting the Dots: A Knowledgeable Path Generator for Commonsense Question Answering [arxiv][project page]

Codes in the below folders are based on Path-Generator-QA.

• Commonsense-Path-Generator Please refer to the following link for installation.

The raw file of domain KG is located at /home/yujin/dot/preprocess_DoKG/raw_data/baby-domain-hiararchy-refined.csv .

If you have any additional needs, please request.

Quick start PG

1. You can dowload trained path generator

   • path generator trained with domain kg checkpoint
   • path generator trained with ConceptNet checkpoint
   • path generator trained with KG integrated domain KG and ConceptNet using naive method checkpoint
   • path generator trained with KG integrated domain KG and ConceptNet using pivoting checkpoint

2. Before run the below code, check the source file path and save path.

   cd learning-generator
   sh run_test_pg.sh

Training Path-Generator for domain specific KG

1. Preprocess Before run the below code, check the source file path and save path. You can download example input file here.

   cd preprocess_DoKG
   sh convert_csv_to_nx.sh

   The outputs of this script are domain kg files (baby_domain_graph.nx / entity_vocab.pkl / relation_freq.pkl / relation_vocab.pkl).

2. Path sampling for path generator Before run the below code, check the input setting.

   • data_dir : dir_path of domain kg files. For example, the save_dir of convert_csv_to_nx.sh above process.
   • output_dir : dir_path for output file
   • graph_file_name : file name of graph_file in data_dir
   • split_dataset : If you set the setting 'split_dataset' as True, split dataset for training by ratio of 0.9:0.05:0.05 will be saved under output_dir.

   cd learning-generator
   sh run_path_sampling_dokg_multi_di_rw_no_hiar.sh 

3. Train path generator using domainKG Before run the below code, check the config file(~/leraning-generator/config/).

   • data_dir : dir_path for training datset. For example, the output_dir of above process.
   • model : name of the LM

   For training a path generator

   sh ./run_dokg_multi_di_rev_no_hiar.sh $gpu_device

Training Path-Generator for integrate CSKG and domain specific KG

1. Preprocess Before run the below code, check the source file path and save path. You can download example input files following links.

   • conceptnet file
   • dokg file
   • integrate list file

   cd preprocess_integrate_kg/integrateKG
   sh integrate_kg.sh

   The outputs of this script are domain kg files (integrate_graph.nx / entity_vocab.pkl / relation_vocab.pkl).

2. For training a path generator

   cd learning-generator
   sh run_path_sampling_inte_pivot.sh

   If you set the setting 'split_dataset' as True, split dataset for training by ratio of 0.9:0.05:0.05 will be saved under output_dir.

3. Train path generator using domainKG

   ./run_inte_pivot.sh $gpu_device

   The output of this script is path generator model.

For training a commonsense qa system

1. Download Data First, you need to download all the necessary data in order to train the model:

   cd commonsense-qa
   bash scripts/download.sh

2. Preprocess To preprocess the data, run:

   python preprocess.py   

3. Using the path generator to connect question-answer entities (Modify ./config/path_generate.config to specify the dataset and gpu device)

   ./scripts/run_generate.sh

   3.1. For domain specific KG (Modify ./config/path_generate_dokg_multi_di.config to specify the dataset and gpu device) bash ./scripts/run_generate_dokg_multi_di.sh

   3.2. For domain specific KG (Modify ./config/path_generate_dokg_multi_di.config to specify the dataset and gpu device) bash ./scripts/run_generate_dokg_multi_di.sh

4. Commonsense QA system training bash scripts/run_main.sh ./config/csqa.config Training process and final evaluation results would be stored in './saved_models/'