qcl 's research -- Detection of Entity Properties in Content Stream

Preprocess for KBA - filter target attributes/properties rapidly.

Motivation

Linguistic knowledge --> world knowledge

• World knowledge varies with time
• How to acquire knowledge from heterogeneous resources to reflect the changes of real wrold is very important (KBA)

Problem

Given a target entity to be tracked, find its (new) related information from heterogeneous resources effectively because large volume of data are created.

• Target entity
  • Different patterns related to the entity type
  • e.g. 歐巴馬 - person type
  • extract all patterns related to person
  • e.g. MS - org type
  • extract pattern related to org.
• Patterns 分類
  • Entites (entities types)
  • Dynamic v.s Static
  • Related information
    • related to some pattern
  • Exact
    • disambiguation
• Evaluation matrics
  • Speed
    • wiki
    • dbpedia
  • acc, prec, recall
  • TREC KBA -> to read KBA for more information.
  • Testing Dataset

Issues

Target

Target entities have different types, different types have different patterns, different patterns related to differnet types.

Note that

• Type
• Pattern
• Feature
• Information

1. Information related to human
   • How many features are related to human ?
   • What kinds of features are related to human ?
   • What kinds of patterns can be used to find the features ?
2. Tell out Dynamic and Static information
   • What kinds of features are dynamic, and what kinds of features are static?
   • In other words, what knowledge is unchanged?
3. Tell out the position of the information
4. Tell out if the information is related to the interesting targets

Efficiency

• Data sturcture and algorithms

• Filtering: (1) and (2)

• Extract: (3) and (4)

• Efficient Filtering, e.g.

  • birth (n patterns related to birth)
  • occupation change (m patterns related to this move)

• If topic model is suitable:

  • birth is a topic, occupation change is a topic,...

• 23.21 mins 24 core 33w

• Spend 1393.446886 seconds

• 1 core spend 24 mins to deal 13750 docs

• 1 core 592docs/mins

• 16 core

• 1432.828743 seconds

• about 3 mins

• 177.767400 s 8 core 6.6w

• 1 core spend 3 min to deal 8250 docs

• 1 core 2750/mins

Effectiveness

• Patterns
  1. entity types
  2. dynamic (new) v.s. static
  3. related information, related to some patterns
  4. exact information, target entites, mention disambiguation

Others (暫時沒想到分法)

• Pattern coverage
• Pattern use

Statistics

• Statistics of samples

• Distribution of features

• Number of features/wiki

• How to sample data from wiki for testing?

• Testing effectiveness and efficient

  • Efficiency: enough documents, number of documents created by human per second, ...

References

• Identifying constant and unique relations by using time-series text
• PATTY: a taxonomy of relational patterns with semantic types
  • http://www.mpi-inf.mpg.de/yago-naga/patty/
  • https://d5gate.ag5.mpi-sb.mpg.de/pattyweb/
• TREC KBA

Tools

• MongoDB
• Stanford parser
• Python
  • RDFLib
  • TextBlob
    • Aptagger
  • NLTK
  • SimpleJson

Dataset Used

• DBpedia
  • Download
    • DBpedia Ontology
    • Raw Infobox Property
    • Raw Infobox Property Definitions
    • Persondata
• Wikipedia dump
  • 20130304
    • enwiki-20130304-pages-articles
• Wiki API,1961414
  • 01-05,58115
  • 06-10,126706
  • 11-15,195114
  • 15-18,262784
  • 19-20,192237
  • 21-22,206117
  • 23,129647
  • 24,142679
  • 25,132130
  • 26,125104
  • 27,390781
• PATTY
  • PATTY Online
• YAGO
  • YAGO Facts