Knowledge Structuring and Applications in Alibaba's Xiaomì Chatbot: From KBQA to EBQA

Alibaba's Xiaomì chatbot family (including Xiaomì, Shop‑Xiaomì, and Cloud‑Xiaomì) serves merchants, enterprises, and government users with an AI‑driven dialogue system that relies on a rich, structured knowledge base composed of FAQs, phrase‑based entries, knowledge graphs, and machine‑read documents.

The knowledge is organized into a two‑level schema: a pattern layer defining types, attributes (with hierarchical sub‑attributes), values (including Text, Key‑Value, and Compound Value Types), and relationships; and an entity layer that instantiates these patterns using "entity‑attribute‑value" and "entity‑relationship‑entity" triples.

To build this schema, the pipeline first extracts domain phrases via n‑gram mining, random‑forest classification, and BERT pruning; then performs concept/entity chunking using a BILSTM+CRF model enriched with Word2Vec, BiLSTM character embeddings, and ELMo/BERT contextual representations; finally, relation extraction is performed with BERT (augmented with entity markers) or GCN, followed by manual verification and CVT design.

For question answering, the system employs Knowledge‑Based QA (KBQA) that converts a user query into a candidate query graph by recognizing entities, attributes, and constraints, ranks the graphs with LambdaMart (GBDT), performs type inference to reduce redundancy, and may recommend clarifying questions when the query is ambiguous.

Event‑Based QA (EBQA) extends KBQA to event‑centric queries, handling WHY/HOW/IF scenarios. It adds an event‑recognition stage (BILSTM+CRF + CNN‑based classification) and attribute recognition (CNN‑BiGRU‑Attention), then follows a similar graph‑ranking and type‑inference process, enabling both factual and hypothetical event answers.

Practical deployments of KBQA/EBQA include Alibaba's large‑promotion robot, merchant robot (Wanxiang), and Taobao Live assistant, which have been stress‑tested during Double‑11 and demonstrated strong performance.

The authors conclude that structured knowledge brings benefits such as accurate answers, reasoning, explainability, and reusability, but also poses challenges like high schema‑construction cost and the need for better tools; future work will focus on reducing schema cost, leveraging multimodal data, and building domain‑specific commonsense graphs.