Below are selected excerpts from the paper:

One of the tasks which may be amenable to computer-assisted intelligence is the generation of summaries or synopses of specific scientific areas. . . .

The generation of the specialty narrative begins with a co-citation cluster derived from a single-link cluster analysis of a citation database. . . . In the example used later in this article, a special 3-year citation database was extracted from the SCI [Science Citation Index] using the interest profile of a single research institute. . . .

To form co-citation clusters, a citation frequency threshold is set to select the most cited documents in the database. . . . The second step is to determine the frequency of co-citation between all pairs of cited documents selected by the threshold. . . .

The subject matter is the study of leukemia viruses, a topic in biomedical science and more specifically cancer virology. . . .

The term “narrative” suggests that we seek a way to transform the structure of the co-citation network into a linear ordering of ideas appropriate to a written text. . . .

The types of trees generated by this algorithm will span the network, but will not always be minimal. . . .

The tree provides us with a linear ordering of the nodes in the network, and a formal structure for the specialty narrative. . . .

With narrative structure established, the next problem is what content should fill it. . . . Rather than use the highly cited documents themselves as a source of this text, we use what the citing authors say about these works when they are cited (that is, the citation contexts). . . .

If it is assumed that most passages express the same concept, but in different words, then it is possible to select that passage which is the most characteristic or typical of the group by virtue of using the most frequently encountered words to describe the idea of the cited work. . . .

The consensus passages for each cited document in the network are arranged in the narrative sequence given by the spanning tree. What is lacking to constitute a coherent narrative are transitions between the passages. . . .

While the citing passages are direct quotes, with only minor editorial changes made for consistency or clarity, the transitional sentences are paraphrases of text which comes between the references in the co-citing text. . . .

The specialty narrative, now completed, is a combination of statements by several individuals from several sources, melded together by common usage, and selected to typify that usage. . . .

We have explored the hypothesis that the thought processes involved in reviewing a field can be modeled by a walk through a co-citation network.

The above summary is not from the Author’s Abstract; it consists of sentences selected from the paper itself. Although these sentences were selected manually, it is possible to automate this process by using clues such as word frequency counts to select representative sentences. The research described extends techniques that have been developed in the context of automatic abstracting to tackle a more ambitious task.

The summary above illustrates both the strengths and weaknesses of this approach. It is possible to identify sentences and passages which express the core ideas of scientific documents. However, such sentence pasting is unlikely to lead to esthetically coherent prose. The approach described is clearly not at a stage where it could replace human reviewers, although it might be used to provide assistance in reviewing. It is also possible that it could provide insights into the review process. Such conclusions would require more comparison with the cognitive behavior of human reviewers than is presented.