Semantic Relation Extraction

• extracting semantics/semantic relations from unstructured text data

An Approach to Lexical-Syntactic Pattern Discovery

1. pick one Semantic Relation at a time
2. find lexico-syntactic patterns that express that SR
3. create a corpus with positive and negative examples
4. learn constraints that discover the SR
5. evaluate performance: precision and recall

Lexical-Syntactic Patterns Expressing Meronymy

Variety of meronymy expressions

Get pattern statistics and select the most frequent ones:

• NP1 of NP2
• NP1’s NP2
• NP1 have NP2

Meronymy Learning Algorithm

• input: positive and negative meronymic examples of concept pairs
• output: semantic constraints on concepts

algorithm:

1. Generalize the training examples:
   • example: {part#sense; whole#sense; target}
   • generalized example: {part#sense, classpart#sense; whole#sense, classwhole#sense; target}
   • e.g. {aria#1, entity#1; opera#1, abstraction#6; Yes}
   • types of examples:
     • positive: {Xclass#sense; Yclass#sense; Yes}
     • negative: {Xclass#sense; Yclass#sense; No}
     • ambiguous: {Xclass#sense; Yclass#sense; Yes/No}
2. Learn constraints for the unambiguous examples (positive and negative)
3. Specialize the ambiguous examples:
   1. based on the IS-A information provided by WordNet
   2. initially, each semantic class represented the root of one of the noun hierarchies in WordNet
   3. Go down in the hierarchy and specialize the semantic classes that correspond to the part, respectively whole
   4. repeat the specialization until no more ambiguous examples in the training corpus