Enriching Document Representation with the Deviations of Word Co-occurrence Frequencies

摘要

Recent strategies have been proposed to reveal the semantic relatedness between documents by enriching a document with the relatedness of all the words in the given document collection to the document. By restricting the relatedness to the expected frequencies that each word will occur in the document, the traditional weighted sum of word vectors is proved to give the upper bounds of the expected frequencies. Duplicate counts usually exist during the sum of the word vectors, which weaken the discriminativeness of the enriched document vectors. The strategy which gives the lower bounds of the expected frequencies is also obtained by keeping the maximum values of the word vectors on each dimension. Together with the lower bounds and the deviations of word co-occurrence frequencies, a novel method is proposed to remove the duplicate counts existing in the upper bounds. As a result, the proposed method smooths the generated document vectors better than the weighted sum strategy. Substantial experiments verify that the document clustering incorporated with the proposed method achieves a significant performance improvement compared with the existing strategies.