Optimizing Knowledge Bias: an Adaptive Correction Method for Open Knowledge Graph Construction

Knowledge graph construction (KGC) aims to extract useful information from text and organize it into structured knowledge graphs (KGs). Some recent methods have utilized the powerful generative abilities of large language models (LLMs) to overcome the limitations of generalization and labor costs in traditional methods. As a result, they have achieved success on small, domain-specific datasets, but still struggle with the challenges posed by the chaotic terms and false facts of open-domain text. The main reason is their lack of ability to refine further knowledge, leading to the accumulation of knowledge biases. To address this issue, we propose an adaptive correction method for open knowledge graph construction, namely KG-OBS, which aims to mitigate knowledge bias by both standardizing knowledge schemas and correcting factual errors. Specifically, we first design a knowledge perceiver to quickly extract knowledge from the text. Next, a knowledge canonicalizer standardizes the schemas of extracted knowledge. During this process, adaptive schema alignment and expansion are achieved using a cluster centroid approximation strategy. Finally, we explore a counterfactual-based knowledge corrector, enabling the model to purify knowledge and reduce factual errors. Additionally, the purified knowledge is encoded and stored in a knowledge retainer for unified management. Experimental results show that KG-OBS can extract high-quality knowledge without training across three KGC benchmarks. Compared with previous works, it not only adaptively expands schema information but also automatically corrects errors to improve KG quality.