Knowledge Structuring for Intelligent Customer Service Upgrade: Alibaba's Knowledge Graph QA Approach

In this presentation, Qiu Likun, an Alibaba algorithm expert, introduces the concept of knowledge structuring to upgrade intelligent customer service, focusing on the transition from traditional FAQ‑based systems to knowledge‑graph‑driven structured QA.

Key challenges include massive knowledge management overhead caused by thousands of standard questions and their numerous paraphrases, and language understanding difficulties such as ambiguity, complexity, and poor cross‑domain reuse.

Solution framework consists of three parts:

Building a structured knowledge graph from unstructured or semi‑structured business documents.

Parsing user queries into structured semantic expressions (Semantic Parsing).

Generating structured answers instead of plain text.

The knowledge‑graph pipeline extracts entities, attributes, and values, forming triples like (Shanghai, population, 24.197 million) and extends them to compound value types. Semantic parsing evolves from simple classification to advanced KAMR (Knowledge‑driven Abstract Meaning Representation) and the KAMR Parser, which handle entity recognition, intent classification, constraint binding, and dependency parsing using models such as BLSTM‑CRF, multi‑factor intent classification, and biaffine dependency parsing.

KAMR ontology defines predicates, operators, types, and properties, enabling cross‑domain queries, elastic granularity, and complex semantics. The multi‑factor intent model decomposes intents into domain, predicate, and target, allowing efficient data reuse and balanced training data.

Benefits of knowledge structuring :

High‑efficiency reuse: a single schema can serve hundreds of entities, reducing knowledge items from thousands to a few hundred.

Precise understanding: semantic parsing identifies entities, attributes, and constraints, delivering concise, accurate answers.

Fine‑grained management: structured schemas simplify addition, deletion, and modification of knowledge.

Reasoning and computation: structured triples enable dynamic calculations, such as deriving company age from founding year.

The overall architecture combines an algorithm layer (KAMR, parsers) with a knowledge layer (KG editing platform, schema editor, entity editor) to provide an end‑to‑end structured QA platform for enterprise customers.

In summary, the report covers the demand analysis, the three‑step knowledge structuring approach, the KAMR‑based solution, and the concrete gains in reuse, understanding, and management for intelligent customer service systems.