Presentation Information
[B-14-19]A Hybrid Inference Architecture of Generative AI and Business Rules for PL Event Extraction
〇Yuko Hirata1, Chihiro Hara2 (1. DOCOMO Technology, Inc., 2. NTT DOCOMO, INC.)
Keywords:
Generative AI,Hybrid Inference Architecture,Text Analysis,Business Rules,Quality Management
NTT DOCOMO, INC. analyzes mobile device failure trends to improve product quality. Extracting Product Liability (PL) events from free-text fault memos relies on manual inspections, causing high workloads and inconsistent quality.To automate this, we tested a Large Language Model (LLM). However, using a single prompt for complex conditions caused conflicts and hallucinations. Moreover, an LLM alone cannot consider essential structured meta-information (e.g., sales dates) or causality, limiting practical accuracy.This paper proposes "A Hybrid Inference Architecture of Generative AI and Business Rules."Aspect-Decomposed LLM Inference: Prompts are divided and executed as binary classification modules for each aspect (e.g., heat) to avoid condition conflicts.System Post-Control: Logical operations are applied to LLM results using attribute data and causal rules to resolve judgment conflicts.Applied to a production environment, this method reduced false positives by up to 62.5% while suppressing false negatives for major events. Retaining intermediate data ensures traceability, realizing a highly practical system.