Awesome

🐐 ALCUNA: Large Language Models Meet New Knowledge

Paper link: https://arxiv.org/abs/2310.14820v1

KnowGen: Entity Knowledge Generation Method

Demonstration of ALCUNA

Algorithm

Data Structure

Everyone can get our dataset from here.

Meta Data

Each line in dataset/meta_data.jsonl corresponds to the metadata information of one artificial entity:

• artificial_entity: the information of artificial entity
  • name: the name of artificial entity
  • id: the negative of parent entity's id
  • rank: rank in the biological taxonomy
  • property: a list of property structure (mentioned later).
• parent_entity: the information of parent entity.
• difference: the differences made for generating the properties of the artificial entity.
  • extension: a list of property structure of artificial entity from other entities.
  • variation: a list of tuples (old_property, new_property), old_property refers to the source of the created property (new_property) of artificial entity.
  • heredity: a list of property structure of artificial entity inherited from parent entity.
  • dropout: a list of property structure of parent entity not inherited to artificial entity.

The property structure is structured as follow:

• name: the name of current property.
• type: the type of current property. One of ['attribute', 'relation'].
• values: a list of valid values of current property.

Question Dataset

dataset/id2question.json contains a dict mapping from artificial entity id to its corresponding questions. Each question is structured as following,

• question
• answers: a list of all valid answers.
• form: the form of the question. One of ['boolean' 'fill-in-blank', 'multi-choice'].
• type: the subset to which the question belongs. One of [ 'Knowledge Understanding', 'Knowledge Differentiation', 'Knowledge Association']
• meta_data:
  • related_property: the property of the artificial entity related to the current question. If difference is 'variation', it is a tuple of (old_property, new_property). For other cases, it is a property structure.
  • difference: the difference type of related_property. One of ['extension', 'variation', 'heredity', 'dropout']
  • hop_triplets (optional): the chain of relation triplets corresponding to the multi-hop question.Only available for Knowledge Association dataset.

Citations

Please cite the paper and star this repo if you use ALCUNA (or KnowGen) and find it interesting/useful, thanks!

@misc{yin2023alcuna,
      title={ALCUNA: Large Language Models Meet New Knowledge}, 
      author={Xunjian Yin and Baizhou Huang and Xiaojun Wan},
      year={2023},
      eprint={2310.14820},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}