KnowledgeDA

KnowledgeDA Framework

Requirements

• Python 3.6
• torch 1.8
• cuda 1.11
• nltk 3.6
• jieba 0.42.1
• synonyms 3.16.0
• scikit-learn 0.24.1
• ordered_set 4.0.2

Quick Start

Data & PLMs Preparation

• Data: unzip the files in data and KG

• PLMs: download PLMs from https://huggingface.co/ and put them into PLMs

  • BERT_base_chinese: https://huggingface.co/bert-base-chinese
  • BERT_base: https://huggingface.co/bert-base-uncased

• After unzipping, the file directory is as follows

.
├── Augment
│   ├── augment_kg_ch.py
│   ├── augment_kg_en.py
│   └── utils
│       ├── cluster_ch.py
│       ├── cluster_en.py
│       ├── entity_mask_ch.py
│       ├── entity_mask_en.py
│       ├── file_handler.py
│       ├── load_data_ch.py
│       └── load_data_en.py
├── Classify
│   ├── aug_quality_assessment.py
│   ├── bert_classification.py
│   ├── BERT_KnowledgeDA.py
│   ├── load_data.py
│   ├── pytorch_pretrained
│   ├── result
│   └── save_dict
│
├── data
│   ├── CMID
│   │   ├── dev.txt
│   │   ├── test.txt
│   │   └── train.txt
│   └── SO-PLC
│       ├── dev.txt
│       ├── test.txt
│       └── train.txt
├── KG
│   ├── CMedicalKG
│   │   ├── entities_dict.txt
│   │   ├── entities_list.txt
│   │   └── triples.txt
│   └── TagKG
│       ├── entities_dict.txt
│       ├── entities_list.txt
│       └── triples.csv
├── PLMs
│   ├── BERT_base
│   └── BERT_base_chinese
├── run_service.py

Run Service

• Example 1: CMID (Chinese)

python run_service.py --dataname CMID --domain healthcare --lan ch --plm_path BERT_base_chinese

• Example 2: SO-PLC (English)

python run_service.py --dataname SO-PLC --domain software --lan en --plm_path BERT_base

• Attention

  • the fine-tuned model will be saved in Classify/save_dict

  • the result will be store as the classification report in Classify/result, here we give an example of the result on CMID

                    precision    recall  f1-score   support
    
                 0     0.8061    0.8421    0.8237       627
                 1     0.7544    0.7681    0.7612       276
                 2     0.5176    0.5665    0.5410       233
                 3     0.6571    0.2556    0.3680        90
    
          accuracy                         0.7300      1226
         macro avg     0.6838    0.6081    0.6235      1226
      weighted avg     0.7287    0.7300    0.7225      1226
    

Extension to other task data

If your task is related to healthcare (Chinese) or software (English), you can directly use our built-in KG; otherwise, you can also use your own KG:

• Put your task data into data

  • include 3 txt files: train, dev and test datasets
  • each row in train.txt and dev.txt should be like: [id]\t[label]\t[text]
  • can refer to the dataset CMID and SO-PLC

• put your domain KG into KG

  • include 2 main files: triples and the dictionary of entity category
  • can refer to the given KG format