Fast mining of interesting phrases from subsets of text corpora

Abstract

We address the problem of mining interesting phrases from subsets of a text corpus where the subset is specified using a set of features such as keywords that form a query. Previous algorithms for the problem have proposed solutions that involve sifting through a phrase dictionary based index or a document-based index where the solution is linear in either the phrase dictionary size or the size of the document subset. We propose the usage of an independence assumption between query keywords given the top correlated phrases, wherein the pre-processing could be reduced to discovering phrases from among the top phrases per each feature in the query. We then outline an indexing mechanism where per-keyword phrase lists are stored either in disk or memory, so that popular aggregation algorithms such as No Random Access and Sort-merge Join may be adapted to do the scoring at real-time to identify the top interesting phrases. Though such an approach is expected to be approximate, we empirically illustrate that very high accuracies (of over 90%) are achieved against the results of exact algorithms. Due to the simplified list-aggregation, we are also able to provide response times that are orders of magnitude better than state-of-the-art algorithms. Interestingly, our disk-based approach outperforms the in-memory baselines by up to hundred times and sometimes more, confirming the superiority of the proposed method.