Baidu Event Graph: Technology, Construction, and Applications

Overview

Baidu has built a million‑scale event graph that updates within minutes and powers search, information streams, and downstream industry services. The article introduces the motivation, core concepts, and the overall architecture of the system.

Event Graph Concepts

An event graph models events, their attributes, and inter‑event relations (temporal, causal, hierarchical). Unlike traditional entity‑centric knowledge graphs, events capture dynamic changes and can be linked to entity graphs.

Technology Overview

The system consists of four layers: data, construction, cognition, and application. The focus is on the construction layer, which includes event detection, representation, extraction, and relation mining.

Event Construction Pipeline

Event detection: fast, accurate, and comprehensive identification of events from news streams using a multi‑task learning framework.

Event representation: automatic discovery of event types and roles from unstructured text, followed by open‑type mining and human verification.

Event extraction: structured extraction of event elements (trigger, type, arguments) via multi‑attribute joint models, reading‑comprehension‑based custom argument extraction, and semantic‑role extraction.

Relation mining: extraction of causal, temporal, co‑reference, and hierarchical relations, with special emphasis on causal reasoning.

Key Techniques

Multi‑task learning to share knowledge among detection models, achieving minute‑level latency and >90% accuracy.

Use of ERNIE pre‑trained language model combined with Bi‑LSTM+CRF for joint attribute extraction.

Reading‑comprehension‑style QA for zero‑shot custom argument extraction.

Semantic‑role abstraction to decouple event types from role vocabularies, enabling efficient multi‑pointer extraction.

Applications

Multi‑dimensional hotspot discovery across industries and regions with minute‑level freshness.

Event timeline/trace generation for users to understand causes and consequences.

Resource association: linking news, images, videos, and other media to events to aid content creation.

Causal reasoning: building a causal graph (≈2 million nodes/edges) for real‑time inference, currently explored in finance.

Dataset and Evaluation

Baidu released the DuEE dataset, the largest Chinese event extraction benchmark, and provides an open leaderboard for community evaluation.

Summary and Outlook

The article concludes with a recap of the event graph’s value, technical achievements, and future directions: enhancing construction capabilities, expanding application scenarios, and deepening industry‑specific explorations.