Relation extraction as a classification problem is formulated as below:

  • r : relation type
  • w(i, j) : span of first argument

  • w(m, n) : span of second argument (might be before, after, or overlap of span of first argument)

Three ways to compute the scoring function

  1. Feature-based classification
  2. Kernel
  3. Neural relation extraction


The scoring function in a feature-based classifier is shown below:

  • Theta : vector of weights
  • f(.) : vector of features

Using pattern-based approach, we can compute several features:

  • Local features of first and second argument of text
    • Strings themselves
    • Are they entities? If so, which type?
    • Each string’s syntactic head
  • Features of span between two arguments
    • Length of span
    • Specific words that appear in span
    • Wordnet synsets for synonyms
  • Features of syntactic relationship between two arguments
    • For example, the dependency path between the arguments. See example below.



Data Scientist

Leave a Reply