Knowledge Organization

Our brain tries to connect words together based on their relationships.

Taxonomy

This connects words based on specificity. For example vegetable can be linked to greens which can be linked to spinach.

Meronymy

This connects words based on a predefined relationship.

For example:

Semantic Network

This connects words based a "one to two word" description.

For example:

• Cat has fur
• Cat is a mammal
• Fish lives in water

Semantics

Definitional

Words can also be connected based on definitions. They can have both taxonomic and meronymic parts.

Predicates / Attributes

A large knowledge base created from the “infobox” sections of wikipedia. Has become the largest knowledge base in the Semantic Web. The structure is in triple form: predicate(subject, object).

Distributional

you know a word by the company it keeps

Trying to figure out a word based on the words around it.

For example:

• I did X today
  • X is probably an activity
• I ate X today
  • X is some sort of food

Overtime and several examples, the distributional semantics get better since the examples work as weights in a neural net.

Semantic Similarity

Definitional Semantic

1. Explicit synonymy
2. Similarity of definitions (comparing the text)
3. Proximity in Taxonomy
   • hypernym: dog/animal
   • co-hyponyms: dog/cat (both having animal as a hypernym)

Predicate Semantic

1. shared predicates between concepts/entities

Distributional

1. vector similarity (cosine)
2. 2d map

Entity Linking

Ambiguity in named entity recognition

• confusion with general language:
  • I have an account at Tangerine
  • I ate a tangerine for breakfast
• confusion between types:
  • I visited Paris last year.
  • Paris Hilton is in the news today.

Ambiguity within a type

• confusion within the same type - entity Linking
  • my friend learned French before going to Paris
    • confusion with Paris and Paris, Texas
  • Michael Jordon was invited speaker at the ML conference.
    • confusion with basketball player and ML researcher

Algorithms

Use Bag of words match algorithm to find similarities with corpus for ambiguous words.

build a bag-of-words BOW_c to contain the words in the context of occurrence of the occurrence of the ambiguous word:
for each possible sense s=1 .. S do:
    Build a bag-of-words BOW_s, to contain the words in the definition of sense s
Assign MaxOverlap to 0
Assign BestSense to null
for each possible sense s=1..S do:
    Measure overlap = Overlap (BOW_c, BOW_s)
    if overlap > MaxOverlap:
        MaxOverlap = overlap
        BestSense = s

Collective Disambiguation referring to the corpus and linking words to entities