Semi-Supervised Graph-Ranking for Text Retrieval

摘要

Much work has been done on supervised ranking for information retrieval, where the goal is to rank all searched documents in a known repository with many labeled query-document pairs. Unfortunately, the labeled pairs are lack because human labeling is often expensive, difficult and time consuming. To address this issue, we employ graph to represent pairwise relationships among the labeled and unlabeled documents, in order that the ranking score can be propagated to their neighbors. Our main contribution in this paper is to propose a semi-supervised ranking method based on graph-ranking and different weighting schemas. Experimental results show that our method called SSG-Rank on 20-newsgroups dataset outperforms supervised ranking (Ranking SVM and PRank) and unsupervised graph ranking significantly.