Event Graphs in Intelligent Customer Service: Concepts, Applications, and System Architecture

1. What is an Event Graph? An event graph (or event‑centric knowledge graph) models events, their attributes, and inter‑event relationships as nodes and edges, extending traditional static knowledge graphs to capture dynamic, time‑bound information such as user actions on e‑commerce platforms.

2. Common Applications Event graphs enable (a) event timeline reconstruction, (b) real‑time hotspot discovery, and (c) downstream event prediction and reasoning. Examples include tracing order‑placement and refund events in a food‑delivery scenario and detecting emerging topics in news streams.

3. Event Graph Construction for Customer Service The workflow consists of (i) Event Schema Induction – defining a schema of event types (e.g., 办理, 现象, 咨询, 投诉) and their trigger predicates; (ii) Real‑time event tagging – extracting key‑phrase candidates from conversation sessions, generating features (position, role, entropy, etc.), and scoring them with a model; (iii) Instance building – mapping identified events to the schema, clustering new candidates, and updating the schema iteratively.

4. Real‑time Processing Framework Data ingestion gathers streaming conversation, review, and order data via Kafka/Binlog and static reference tables from Hive. A Flink+Spark engine applies the event‑tagging model, stores enriched records in Doris (wide tables) and Elasticsearch for text search, and exposes APIs for downstream consumption.

5. Applications in Intelligent Customer Service Event Script Prediction models narrative event chains (e.g., restaurant visit) to anticipate next user actions, forming Taskflow trees for task‑oriented bots. Multi‑turn Topic Guidance uses the graph to steer dialogues toward relevant sub‑issues (e.g., membership cancellation reasons). Business Decision Support aggregates event data across merchants, products, and regions to surface safety incidents, monitor trends, and inform operational interventions.

6. Outlook The current implementation is a prototype limited to a single business line; future work aims to enrich the event graph, expand to more domains, and further validate its impact on QA and decision‑making.