知识图谱问答

2024年后，更多参考graphRAG设计

1. requirements

Functional Requirements:

• Natural language question input

• Accurate answers based on knowledge graph information

• Support for complex multi-hop reasoning

• Handle different question types (factoid, relationship, comparison)

Non-Functional Requirements:

• Low latency (<500ms response time)

• High availability (99.9%)

• Scalability to handle large knowledge graphs

• Support multiple languages

• Privacy and security compliance

2. ML task & pipeline

a) Question Understanding:

• Question type classification

• Entity recognition and linking

• Relation extraction

• Query intent classification

b) Graph Processing:

• Graph embedding generation

• Subgraph retrieval

• Path ranking

c) Answer Generation:

• Answer extraction/generation

• Confidence scoring

• Evidence compilation

3. data collection

Sources:

• Knowledge Graphs:

  • Wikidata

  • DBpedia

  • Domain-specific KGs

  • Company internal KGs

• Training Data:

  • WebQuestions

  • ComplexWebQuestions

  • LC-QuAD 2.0

  • KQA Pro

  • MetaQA

Data Processing:

• Entity normalization

• Relation alignment

• Graph completion

• Question-answer pair generation

4. feature

Question Features:

• BERT/RoBERTa embeddings

• Dependency parsing features

• Named entity mentions

• Question type indicators

Graph Features:

• Node embeddings (TransE, RotatE)

• Structural features

• Path features

• Subgraph features

5. model

6. evaluation

Metrics:

• Accuracy

• F1 Score

• Hits@K

• MRR (Mean Reciprocal Rank)

• Path validity

• Answer completeness

• Reasoning correctness

Testing Approaches:

• Unit tests for each component

• Integration tests

• A/B testing

• Human evaluation

• Adversarial testing

7. deployment & serving

Infrastructure:

• Containerization with Docker

• Kubernetes for orchestration

• GPU support for inference

• Load balancing

• Auto-scaling

8. monitor & maintenance

Monitoring:

• Model performance metrics

• System health metrics

• Error rates and types

• Latency distribution

• Resource utilization

reference

• https://github.com/shawnh2/QA-CivilAviationKG

• 美团知识图谱问答技术实践与探索

• 检索式对话系统在美团客服场景的探索与实践

• 阿里小蜜：知识结构化推动智能客服升级

• 阿里实时语音与智能对话

• 智能机器人在滴滴出行场景的技术探索

• QQ浏览器智能问答技术探索实践

• 达摩院基于元学习的对话系统

• 大话知识图谱-意图识别和槽位填充