Clustering retrieved documents is a typical post-retrieval processing technique used to present an organized result set, not simply a ranked list, to the user, in order to reduce the cognitive burden of going through a large number of returned results. Some commercial search engines, such as Vivisimo, have implemented this strategy fairly successfully. However, studies show that the benefit of clustering is undermined by a poor visual connection between the clusters and the document content. By providing extra indicative extracts, covering both the similarities and particularities of each document contributing to the specific cluster, the authors of this paper take a further step toward improving retrieved documents’ organization and accessibility.

With the assumption that each document consists of several subtopics, the authors first use the TextTiling algorithm to segment the documents. K-means variants are then used to cluster the text segment, and a sentence extraction-based multidocument summary is generated for each cluster, to cover common aspects using surface level information (for example, locations, headings, tf*idf values, and so on). Finally, a single summary is generated for each document indicating its originality. Commonality detection is relatively easier than difference identification, in that, for the latter problem, it is even harder to balance originality and relevance. Text segment alignment is also necessary if multiple aspects are addressed for the same topic. Regretfully, this paper does not offer a detailed solution for difference and originality detection.

The authors used both objective (instance precision/recall) and subjective (questionnaires) methods to evaluate their system’s effectiveness. The objective evaluation shows that commonality summarization helps reduce the reading load by 20 to 30 percent, and it is not a surprise to see that the difference summary does not significantly help. The usability problem still remains in this approach, based on a subjective evaluation by users that it is hard to use the new system.