General Terms

The authors attempt to automatically summarize an electronic brainstorming session (EBS) through the selection of term clusters that are representative of the important topics raised in the EBS. An EBS is structured as a collection of textual comments made by the human participants. The procedures used are essentially statistical, with the exception of a preliminary word exclusion and normalization through the stop-wording of 1000 common function words and "pure verbs" such as "calculate" (although verb forms are not typically used in topic terms, verb elimination is not otherwise justified) and word stemming (neither the rules for and utility of dropping 22 suffixes, nor the comparison with other stemming algorithms, are made clear in detail).

Topic terms are selected from the remaining words and strings of up to three words. Only "common" terms are selected, by eliminating from consideration any term with fewer than four occurrences in the EBS corpus. The "combining weight" of a term j in a comment i is defined as the product of the term's frequency in comment i and logdfj , where dfj is the number of comments containing term j . (As the authors note, this is contrary to the common statistical measures in information retrieval where the inverse document frequency is used to heavily weight terms used more commonly in the document than in the corpus.) An extension of this parameter is used to define the strength of the association of one term with another, thereby establishing what the authors call an asymmetric concept space matrix. Finally, a Hopfield network activation procedure is used to cluster the terms, the strong clusters being taken as concepts most representative of the central EBS topics.

These results were compared against topic lists created by members of the EBS itself as well as by independent human reviewers of the EBS corpus. Despite inspired efforts to use ranking and core term recall and precision results, the considerable variation among the topic lists and judgments made valid, quantified comparisons difficult. To the extent that, in several cases, the authors' system list rated better than the worst of the other five lists compared, one may say that the system shows some promise. Nevertheless, despite an interesting and valiant attempt to solve a difficult problem, the authors' efforts may be best remembered as a useful benchmark indicating the difficulty of the problem and suggesting that more sophisticated and interactive techniques than simple statistical ones will be needed to do an effective job in handling such complex linguistic communication tasks.