SKG-VLA: Scene Knowledge Graphs for Multimodal Complaint Decision Making
A new AI framework, SKG-VLA, models complaint cases as structured scenes using Scene Knowledge Graphs (SKG) to integrate heterogeneous evidence like narratives, screenshots, metadata, and policies. The system organizes entities, evidence, clauses, events, and relations into a unified graph for better decision making. A data synthesis pipeline generates rule-consistent scene descriptions. The approach aims to overcome limitations of shallow classification and template matching in large-scale complaint handling.
Key facts
- SKG-VLA is presented for multimodal complaint decision making.
- It models each case as a structured complaint scene.
- Scene Knowledge Graph (SKG) organizes complaint entities, evidence items, policy clauses, temporal events, transactional states, and action-relevant relations.
- A data synthesis pipeline generates complaint scene descriptions and rule-consistent outputs.
- Existing systems perform shallow classification or template matching over isolated modalities.
- The framework addresses underutilization of explicit scene structure, rule knowledge, and cross-evidence dependencies.
- Decision making relies on heterogeneous evidence including complaint narratives, screenshots, order metadata, historical interactions, and platform policies.
- The core idea is to represent decision-relevant semantics with SKG.
Entities
Institutions
- arXiv