Semantic analysis and representation

The semantic analysis of a word, a sentence or a text, consists in determining its meaning. The notion of meaning is nevertheless ambiguous and can only be precisely defined with respect to a given formalism that is called semantic representation. The current chapter aims at presenting the formalisms developed and used by Japanese laboratories.

Two main families of semantic representations can be observed:

• structured organizations of semantic units, called concepts, usually with the help of graphs;
• non-sorted series of concepts that define a high-dimension vectorial space whose vectors are meanings. This representation enables simple tools of Euclidean geometry to be used, e.g. innerproduct or projections.

• Structured organizations of concepts
  • A concept dictionary at EDR
  • Co-occurrence graphs at ETL
  • Hierarchical clustering of words at ATR
• Non-organized semantic representation
  • A vectorial approach for information extraction at NTT
  • Semantic vectors at Shinshu University
  • A semantic representation using neural networks at ETL
• Mixed approaches
  • Automatic classification of documents at Tokushima University
  • Words similarity measurement at NTT CS Laboratory