SAC-KG - A general knowledge graph construction framework capable of building domain knowledge graphs with over one million nodes.

What is SAC-KG?

SAC-KG is a general framework for automatically constructing domain knowledge graphs based on large-scale language models (LLMs). Composed of three main components—a generator, a validator, and a pruner—it can automatically generate a first-level knowledge graph for a specific domain from a raw domain corpus, ensuring the accuracy of the generated triples. SAC-KG can construct domain knowledge graphs with over one million nodes, achieving an accuracy of 89.32%, which is more than 20% higher than the current state-of-the-art methods. This framework effectively utilizes LLMs as domain experts to generate professional and accurate multi-level knowledge graphs.

Main functions of SAC-KG

• Automated knowledge graph constructionSAC-KG automatically constructs single-layer knowledge graphs for a specific domain from the original domain corpus through its components, and can iteratively construct multi-level knowledge graphs.
• Improve construction accuracyThe SAC-KG validator and pruner work together to ensure accuracy by correcting generation errors and determining whether the newly generated tail needs to iterate on the next level of the knowledge graph. In the experiment, the accuracy reached 89.32%, which is more than 20% higher than the existing methods.
• SpecializationBased on large language models (LLMs) as domain experts, the knowledge graph generated by SAC-KG has a high degree of specialization and can generate triples related to specific domains.
• Control the generation processBy introducing an open knowledge retrieval system and a pruner, SAC-KG can effectively control the generation process, ensuring that the generated triples are correctly formatted and meet domain requirements.
• Large-scale build capabilitySAC-KG can automatically construct domain knowledge graphs on a scale of over one million nodes, demonstrating its advantages in handling large-scale datasets.
• Unsupervised methodsSAC-KG is an unsupervised method that can be applied to any domain with a large unstructured text corpus without labeling the data.
• Consistency assessmentThe high quality and reliability of the knowledge graph generated by SAC-KG were verified by comparing its consistency with GPT-4 and human assessments.

SAC-KG Technical Principles

• GeneratorThe generator is responsible for retrieving relevant information from the original domain corpus and open-source knowledge graphs, feeding them as input to large language models (LLMs) to generate a domain-specific first-level knowledge graph. It consists of two sub-modules:
  • Domain Corpus RetrievalThe domain corpus is segmented into sentences, sorted according to the frequency of entity occurrence, and finally the sorted sentences are concatenated into a fixed-length text and input into LLMs.
  • Open Knowledge RetrievalFor entities in the open-source knowledge graph, relevant triples are provided as examples; for entities not in the open-source knowledge graph, they are segmented and retrieved again; for entities that still do not match, ten triples are randomly selected as suggestions.
• VerifierThe validator's role is to detect and filter out erroneous triples generated by the generator. This involves two steps:
  • Error detectionUsing over 7,000 rules mined from RuleHub, the generated triples are checked for quantity, format, and conflict.
  • Error correctionBased on the detected error type, provide corresponding prompts and prompt LLMs again to generate the correct triples.
• PrunerThe pruning and validator work together to ensure accuracy by determining whether newly generated tails need to iterate on the next level of the knowledge graph, thus enhancing the controllability of the constructed knowledge graph. Based on the open-source knowledge graph DBpedia, a fine-tuned T5 binary classifier model takes as input the tail entity of each correct triplet and outputs "growing" or "pruned," indicating whether the entity continues to generate the next level of the knowledge graph or stops further generation.

SAC-KG project address

• arXiv technical paper:https://arxiv.org/html/2410.02811v1

Application scenarios of SAC-KG

• Construction of professional domain knowledge graphsSAC-KG can be applied in multiple fields such as medicine, biology, and social networks to help build knowledge graphs for specific domains.
• Automation and improved accuracySAC-KG improves the automation and accuracy of knowledge graph construction through an automated construction process, achieving an accuracy of 89.32%, which is more than 20% higher than the current state-of-the-art methods.
• SpecializationSAC-KG utilizes large language models (LLMs) as domain experts to generate domain-specific triples, giving the knowledge graph a high degree of specialization.
• Large-scale data processingSAC-KG can automatically construct domain knowledge graphs on a scale of over one million nodes, demonstrating its advantages in handling large-scale datasets.