- An implementation for Memorizing is Not Enough: Deep Knowledge Injection Through Reasoning (ACL 2025).
- Please contact @Ruoxi Xu (ruoxi2021@iscas.ac.cn) for questions and suggestions.
Moves beyond simple memorization to enable deep knowledge reasoning through automated data augmentation and multi-level evaluation.
- Diversified Data Augmentation
- Multi-level Test Cases Generation
- Recall-level: Basic knowledge recall
- Extraction-level: Knowledge extraction from complex contexts
- Reasoning-level: Logical inference and knowledge application
knowledge-learning-toolkits
├── data_augumentation # Knowledge augmentation module
├── ── augment.py # Data augmentation
├── ── prompt_hub.py # Defines prompts for data generation and functions for parsing returned results
├── data_loader # Data preprocessing module
├── ── load_data.py # Reads and preprocesses data; converts structured and unstructured data into a unified format: [{'idx': 0, 'text': text, 'knowledges': [{'text': text, 'triplet': [subject, relation, object]}]}]
├── evaluator # Evaluation module
├── ── generate_test_cases.py # Automatically constructs test cases
├── trainer # Train module
├── utils # Common utility functions
├── config.yaml # Configuration settings
└── main.py # Main entry script
The toolkit currently accepts two types of data formats:
- For data in the form of triples, it should be stored in a .jsonl file, with each entry in the following format: {'idx': 0, 'triplet': [subject, relation, object]}
- For free-text data, it should also be stored in a .jsonl file, with each entry in the following format: {'idx': 0, 'text': text}
python main.pyIf you find this project helpful, please use the following to cite it:
@article{xu2025memorizing,
title={Memorizing is Not Enough: Deep Knowledge Injection Through Reasoning},
author={Xu, Ruoxi and Ji, Yunjie and Cao, Boxi and Lu, Yaojie and Lin, Hongyu and Han, Xianpei and He, Ben and Sun, Yingfei and Li, Xiangang and Sun, Le},
journal={arXiv preprint arXiv:2504.00472},
year={2025}
}