Automatic structuring and retrieval of large text files

Reviewer: Harold Borko

Salton is a well-known and respected teacher and researcher in the areas of text processing and information storage and retrieval as well as a prolific writer on these topics. This paper, written in collaboration with his students, describes the results of experiments involving the structuring of large bodies of text by linking excerpts containing related subject matter. The experiments utilize many common text processing procedures, some of which have been developed by Salton and his colleagues, plus a number of innovative strategies for text matching and retrieval. The database used in these experiments consists of 25,000 articles contained in the 29 volumes of the <__?__Pub Fmt italic>Funk and <__?__Pub Fmt nolinebreak>Wagnalls<__?__Pub Fmt /nolinebreak> New Encyclopedia<__?__Pub Fmt /italic>, about 65MB of text. Salton states that, to retrieve relevant information from such a large body of text, the use of keywords connected by Boolean operators would give inadequate results. He proposes a vector processing model in which both the stored documents and the search requests are represented by sets (vectors) of terms without Boolean operators. Searching is done by comparing query vectors with the vectors representing the stored documents and retrieving those items found to be similar to the queries. In the reported experiment, an existing encyclopedia article is used as a search request, and the goal is to retrieve related articles. The formula for computing the vector similarity function is given and discussed, and the search results, using a number of different conditions, are displayed in easily readable tables and figures. As Salton points out, “The evaluation of information retrieval or text linking operations is a major unsolved problem” (p.<__?__Pub Fmt interword-space>105). The problem is largely due to the difficulty in obtaining accurate relevance judgments, which are needed to compute recall and precision parameters. Recognizing that one cannot obtain absolute performance measurements, this experiment computes relative performance results for the different search runs in order to judge the effectiveness of the retrieval methodologies. Both the experimental and the evaluation procedures are reasonable and appropriate. No mention is made of the cost and computer processing time required to prepare keyword vectors for the encyclopedia articles and to conduct the search either by a simple matching procedure or by the use of relevance feedback, however. Could these techniques be practically and affordably applied to existing information retrieval databases__?__ At the very least,<__?__Pub Caret> this question should have been considered.