Document Recovery from Bag-of-Word Indices

Motivated by computer privacy issues, we present the novel problem of document recovery from an index: given only a document’s bag-of-words (BOW) vector or other type of index, reconstruct the original ordered document. We investigate a variety of index types, including count-based BOW vectors, stopwords-removed count BOW vectors, indicator BOW vectors, and bigram count vectors. We formulate the problem as hypothesis rescoring using A∗ search with the Google Web 1T 5-gram corpus. Our experiments on five domains indicate that if original documents are short, the documents can be recovered with high accuracy.