Knowledge Graph‑Based Question Answering in Meituan’s Intelligent Interaction Scenarios

The presentation begins with an overview of six IT industry waves, emphasizing the evolution of human‑computer interaction from mouse‑keyboard to voice‑driven AIOT, and introduces the use of knowledge‑graph QA within Meituan’s intelligent interaction platforms.

1. Intelligent Interaction Taxonomy

Retrieval‑style interaction – classic FAQ matching.

Task‑oriented interaction – e.g., booking flights or hotels.

Chat‑style interaction – entertainment‑focused, end‑to‑end deep‑learning systems.

2. Meituan Life‑Service Interactions

Meituan covers catering, entertainment, hotel, and travel. Different domains require different interaction types, and many user intents (e.g., product comparison, ticket booking) are difficult to satisfy with current UI flows.

3. Knowledge‑Graph QA (KBQA) Features

Data preparation – building domain‑specific graphs with expert input.

Data management – easier maintenance due to structured knowledge.

Intent detection – locating sub‑graphs for high‑precision answers.

Question types – supports multi‑hop, constrained reasoning.

Result precision – higher answer accuracy thanks to explicit knowledge.

Multi‑turn capability – handles dialogue rounds effectively.

4. Traditional KB‑QA Approaches

Two main streams:

Semantic‑Parsing‑based KB‑QA: maps a question to a query language (e.g., SPARQL) and executes it directly on the graph.

Information‑Retrieval‑based KB‑QA: extracts core entities, retrieves a candidate sub‑graph, then generates an answer via ranking or template matching.

5. Semantic Parsing Details

Following EMNLP 2013, the pipeline includes mapping words to entities/relations, constructing a syntax tree with link, intersect, and aggregate operations, and classifying the correct tree to produce a query.

6. Information Retrieval Details

Based on arXiv 2016, the pipeline encodes the question with a Bi‑LSTM plus attention, encodes candidate answers (entity, relation, type, context), matches them, and returns the highest‑scoring answer.

7. Empirical Comparison

On the WebQuestions dataset, Semantic‑Parsing + Deep Learning achieves the best F1 (45‑55), while end‑to‑end IR struggles with complex queries.

8. Adapting to Meituan

Because Meituan domains are many and loosely coupled, pure semantic parsing needs abundant annotated data, whereas IR lacks interpretability. The solution combines both: use IR to locate a constrained sub‑graph, then apply semantic parsing for an explainable SPARQL query.

9. Complex Scenario QA

Examples illustrate challenges such as entity explosion (numerous product variants), non‑connected queries (e.g., “spicy restaurant”), and context fusion across multi‑turn dialogues.

10. Three‑Layer Concept Nodes

Product/Standard items – concrete goods with attributes like barcode.

Homogeneous non‑standard items – same ontology, special distinguishing attributes (e.g., brand, grade).

Heterogeneous pure concepts – human‑crafted categories (e.g., consumer groups, business districts).

These concepts are stored using a Freebase‑style CVT model, allowing efficient linking and attribute completion.

11. Benefits of Concept Introduction

Optimized entity linking – reduces ambiguous matches.

Improved information and comparison queries – comparisons can be performed at the concept level instead of enumerating millions of items.

Missing attribute completion – fallback to concept‑level defaults when specific data is absent.

12. Handling Non‑Connected Queries

Attribute propagation and path‑walking (entity‑relation paths and ontology‑based paths) enable answering queries like “spicy restaurant” or “ticket availability at 4 pm”. Offline edge enrichment further reduces online computation.

13. Context Fusion in Multi‑Turn Tasks

The architecture is refactored into three layers: understanding (intent redirection, coreference resolution, ID schema extraction), generation (session slot alignment, function construction), and output (producing query functions and intents instead of raw answers). This decouples KBQA components for seamless integration with task‑oriented dialogue systems.

14. Conclusion

Complex‑scenario QA demands not only algorithmic advances but also robust knowledge construction and tight coupling with multi‑turn interaction systems; the presented three‑layer concept model and hybrid KBQA pipeline address these challenges in Meituan’s real‑world applications.