Doctor Takahashi from ETL laboratories proposed a method to represent semantic contents of Japanese words or sentences with real vectors, thanks to neural networks.
These vectors, called semantic representation vectors (SRV) have all a fixed size and are obtained thanks to recursive auto-associative memories (RAAM) neural networks trained on a corpus, in order to assign similar SRV to similar words or sentences. Moreover the transformation from a sentence to a SRV is reversible so different sentences have different SRVs.
The semantic representation is thus obtained in a vector space whose base vectors are not defined by a list of concepts but automatically adjusted by neural networks.