On the relevance of documents for semantic representation
Laurianne Sitbon, Peter Bruza · 2008
Abstract The subject of this paper is the quality of semantic vector representation with random projection under various conditions. The main effect we are watching is the size of the context in which words are observed. We are also interested in the stability of such representations since they rely on random initialisation. In particular we investigate the possibility of stabilising terms representations through documents representations. The quality of semantic representation was tested by means of synonym finding task using the TOEFL test on the TASA corpus. It was found that small context windows produces the best semantic vectors with 59.4 % of the questions correctly answered. Processing the projection between terms and documents representations several times was found not to improve the stability of the representation. It was also found not to improve the average quality of representations.