The article discusses the construction of a knowledge graph (KG) using fine-tuned large language models (LLMs), emphasizing the importance of data quality. Two significant challenges were identified: the necessity for high-quality training sets and the limitations of LLMs in entity recognition tasks. To overcome these issues, the authors proposed recursively generating training data from inference outputs, which, after manual verification, can be integrated to improve model performance. This approach showcases a method to enhance model effectiveness, particularly for named entity recognition (NER) and relation extraction (RE) tasks.
In our research, we successfully designed a functional material KG by employing fine-tuned large language models (LLMs), assuring traceability throughout the information process.
The main challenge we faced was the need for high-quality training data, which was successfully addressed through automatic recursive data generation from inference results.
#machine-learning #natural-language-processing #data-quality #knowledge-graphs #functional-materials
Collection
[
|
...
]