Improving the quality of the top results of Web search engines is one of the primary research issues in information retrieval. It has become even more important since most Web users usually only look at the top one or two search result pages. Some people even suspect a modern-day Faust-like agreement made with the devil by some young researcher trying to achieve ultimate precision at the top ranks.

The authors of this paper try to achieve something similar without such an agreement by integrating document-, cluster-, and passage-based information for re-ranking search results. In this application, cluster- and passage-based approaches can be seen as expanded and contracted document representations, respectively.

For integration purposes, the authors develop two different methods for re-ranking documents in an initially retrieved set. The first is a probabilistic language model and the other is a discriminative approach that uses language model-based estimates. They perform extensive experiments using several Text Retrieval Conference (TREC) test collections to analyze the merits of the individual methods and their different combinations. They show that integration works.

The paper provides a careful literature survey and a detailed presentation. It needs careful and patient reading. In this regard, some additional intuitive figures and summary tables would have been useful. It is a good paper for researchers.