Leveraging cross-domain knowledge alignment for AI transformation: A contrastive learning approach using problem-solution pairs from patents

초록

Although increasing attention has been paid to organizational factors that enable successful artificial intelligence (AI) transformation, limited guidance exists on how organizations can effectively explore AI knowledge relevant to problems in their specific domains. This process typically relies on intensive cross-domain collaboration among experts, which is time-consuming, labor-intensive, and susceptible to siloed thinking. As a remedy, we propose a systematic analytical framework for identifying AI transformation opportunities by aligning problemsolving knowledge across target and AI domains. First, we extract key problems and solutions from technical documents, enabling a structured representation of knowledge in each domain. Next, we use contrastive learning to model implicit associations between problems and solutions from different domains based on their industrial contexts and underlying technical functions, thereby constructing an industrial context-technical function landscape as a joint embedding space. Within this space, we identify AI-domain problems and solutions that are proximate to the target problem, where proximity reflects both contextual relatedness and technical feasibility. A case study using 24,214 business method patents and 35,241 AI patents confirms that the proposed approach effectively identifies technically relevant and novel AI solutions for target problems, significantly outperforming random retrieval and KorPatBERT, a state-of-the-art patent-domain language model baseline.

키워드