Presentation Information
[BI-3-02]Network Diagram Structuring via Cross-Modal Analysis of Markup Language and Images Using Generative AI
〇Toshiyuki Adachi1, Yuji Oishi1, Junnosuke Wakai1, Hitoshi Watanabe1 (1. Hitachi, Ltd.)
Keywords:
Network,Telecommunication,Unstructured data,Diagram,Generative AI
Generative AI is expected to advance network operations, but field network diagrams exist as unstructured data, and accurate structuring remains a challenge. Such diagrams are often drawn with tools such as PowerPoint, whose internal representation is a markup language such as XML. We propose structuring them by cross-modal analysis of this markup and the rendered image. The two modalities are complementary: where lines cross or crowd, image analysis misreads connections that the markup preserves, while conventional expressions the markup cannot capture, such as plain-line connections and grouped symbols, are read from the image. A generative AI agent referencing both the image and the XML selects the suitable analysis, resolves ambiguous endpoints by isolating and re-rendering shapes with color coding, and outputs JSON-schema data. On the public dataset JPNM48 (48 nodes, 82 edges), we compared it with a conventional method using the image only. Node F1 was 1.000 for both; edge F1 was 0.966 (proposed) versus 0.806 (conventional). The proposed method greatly reduced false positives and negatives, a Welch's t-test confirming a statistically significant advantage.