[4Yin-A] Poster 6

[4Yin-A-55] Automated Crosswalk Toward Greater Interoperability in AI Safety

〇Takayuki Semitsu^1,2, Naoto Kiribuchi^1,2, Kengo Zenitani^1,2, Hikaru Matsuoka^3,1,2 (1. Japan AI Safety Institute, 2. Information-technology Promotion Agency, 3. RIKEN Center for Advanced Intelligence Project)

Keywords:

Crosswalk,AI Safety,Taxonomy,Interoperability

Discussions on AI safety span a wide range of materials, from policy documents and official statements to datasets and benchmarks. However, since many of these related resources are unstructured, it is difficult to objectively compare equivalences and differences across documents. This lack of structured comparison methods poses a challenge for ensuring interoperability among document-based international initiatives. In this study, we propose a method to automatically generate a “crosswalk” that organizes commonalities and differences between a pair of unstructured documents based on a predefined taxonomy. Specifically, we present two applications: a “policy crosswalk” for policy-related documents and an “evaluation-resource crosswalk” for datasets used in AI safety evaluation. On the technical side, we develop an analysis pipeline that uses large language models (LLMs) to automatically interpret and classify documents. We apply the proposed method to representative international AI safety resources and examine its feasibility.