KDD2014-KG Tutorial

知识图谱

Knowledge Graphs

* The Breadth Hypothesis:" To behave intelligently in unexpected situations, **an agent must be capable of falling back on increasingly general knowledge.**"
* Why Graphs:实体及它们之间的关系客观地广泛存在，图能方便地刻画这种事实。
* 实例：Freebase, yago, DBpedia, Facebook's Entity Graph, MS's Satori, Google's Knowledge Graph, OpenIE(Reverb, OLLIE), ConceptNet

Research Challenges

* How many facts are there? 全空间有多大？
* How many of them can we represent? 已知的事实有多少？
* How much the boundaries of our current knowledge limit what we can learn?当前知识的边界对我们有怎样的限制？
* How many facts can be potentially extracted from text? 有多少事实已经被记录下来了，并且在将来可能会被机器学习得到？
* Error in extraction, e.g., parsing errors, overly general patterns
* noisy/unreliable/conflicting info
* disparity of opinion, e.g., Who invented the radio?
* Quantifying completeness of converage
* fictional contexts, e.g.,abraham_lincoln is a vampire_hunter.
* outright spam 彻头彻尾的垃圾。。
* shouLd we trust all sources equally?
* negative examples
* entity resolution/deduplication
* commonsense knowledge
* how to jointly acquire knowledge from different data sources like text, video, sounds, images & sensor collections.

Methods and Techniques

• relation extraction
  
  • Supervised models
  • Semi-supervised models
  • Distant supervision
• entity resolution
  
  • single entity methods
  • relational methods
• link prediction
  
  • rule-based methods
  • probabilistic models
  • factorization methods
  • embedding models
• entity classification
• group/expert detection
• ontology alignment
• object ranking

relation extraction

• Single-entity entity resolution
  entity resolution without using the relational context of entities
• distances/similarities based method
  Edit distance(Levenshtein, etc)
  Set similarity(TF-IDF, etc)
  Alignment-based
  Numeric distance between values
  Phonetic similarity
  Equality on a boolean predicate
  Translation-based
  Domain-specific

重点的意思

1. Structured Knowledge Representation means both entities and relations come from a fixed lexicon (rather than strings).
2. The world changes, but we don't retract facts.
3. Commonsense infomation is too obvious to collect.
4. IE = Infomation Extraction
5. open IE = extracting unstrctured facts from unstrctured sources (text).
6. OLLIE = Open Language Learning for Info Extraction.
7. KDD104页，single entity resolution & relational entity resolution有什么区别
8. entity resolution = 实体识别 = 实体统一
9.