- Resource Description Framework (RDF)
- Resource: object that has URI identification, such as web page, images, videos
- Description: attributes, features and relations among resources
- Framework: description model, language and syntax
- Basic unit, triple pattern: Subject(主语) -- predicate(谓语) -- Object(宾语)
- Hop: path from source entity to target entity
- SPARQL: query language for RDF
- Variable in RDF starts with '?' or '$'
- Compared with deep learning, knowledge graph provides interpretable.
- 1. Data
- 1.1 Structured data (graph mapping)
- E.g., relational database (i.e., Database2Rdf), open kg (i.e., linked data, graph mapping 图映射)
- 1.2 Semi-structured data (wrapper)
- E.g., web page, web table, Wikipedia infobox
- 1.3 Unstructured data (information extraction, often in closed domin)
- E.g., natural language, images, video
- 1.1 Structured data (graph mapping)
- 2. Knowledge extraction from 1.3 Unstructured data
- 2.1 Entity extraction
- Value/number detection and recognition
- Running example
- Entity linking
- Definition: Find the entity (i.e., entity mention 实体指称项) in text and linking it to existing knowledge graph
- Entity Disambiguation 实体消歧, page rank
- Co-reference Resolution (CR) 共指消解
- 2.2 Relation extraction
- Input: unstructured text, a group of entities. Output: a group of triplets, e.g., (First Entity, Second Entity, Relation Type)
- Methods [1, 2]
- Pattern / rule matching
- Trigger word pattern
- Dependency parsing pattern, verb is trigger word. Running example
- Supervised method
- Running example
- Two classifier
- Yes or no classifier determine if there is a relation
- Relation classifier determine the exact relation
- Semi-supervised method
- Bootstrapping
- Example [1].
- Distant supervision
- Input: unstructured text, database contains known entity relations. Output; a set of labeled data
- Combination of bootstrapping and supervised. Cannot find new relationship. Example [1].
- Deep model: PCNN
- Bootstrapping
- Pattern / rule matching
- 2.3 Event extraction
- Trigger word 触发词
- Time 时间
- Location 地点
- Event detection and tracking
- 2.1 Entity extraction
- ** 3.Knowledge fusion**
- Entity alignment
- Similar description
- Similar attribute - value
- Similar neighbor entities
- Schema matching
- Instance matching
- Entity alignment
- Knowledge representation learning
- Convert entity and relationship into vectors
- Application: link prediction (given S and P, predict O) or relation prediction, knowledge reasoning
- Translation based Methods
- TransE: head + relation = tail
- Drawbacks
- cannot process one-vs-multiple (一对多), multiple-vs-one (多对一) or multiple-vs-multiple (多对多) relationship
- cannot process symmetric relationship
- Drawbacks
- TransH
- TransR
- TransE: head + relation = tail
- Protege: edit ontology by hand, visualization, and reasoning
- DeepDive: relation extraction, tutorial
- Nebula, used by MeiTuan
- Question answering
- How to convert natural language to query language ?
- Searching
- Recommender system
- Event prediction
- Knowledge reasoning
- Financial
- OpenKG.CN publishes most recent report.
- CN-DBPedia extract structured information from Baidu Baike.
- Zhishi.me ensembles Baidu Baike, Hudong Baike and Chinese Wikipedia. Dump data
- Agriculture Knowledge Graph
- Knowledge graph demo
- KBQA based on ES