return articles about semiconductors, food of various kinds, small pieces of wood or stone, golf and tennis shots, poker games, people named Chip, etc.

The other side of the problem is that we miss relevant information (and this is much harder to know about!). In controlled experimental tests, searches routinely miss 50-80% of the known relevant materials. There is tremendous diversity in the words that people use to describe the same idea or concept (synonymy). We have found that the probability that two people assign the same main content descriptor to an object is 10-20%, depending some on the task (Furnas et al., 1987). If an author uses one word to describe an idea and a searcher another word to describe the same idea, relevant material will be missed. Even a simple concrete object like a "viewgraph" is also called a "transparency", "overhead", "slide", "foil", and so on.

Another way to think about these retrieval problems is that word-matching methods treat words as if they are uncorrelated or independent. A query about "automobiles" is no more likely to retrieve an article about "cars" than one "elephants" if neither article contains precisely the word automobile. This property is clearly untrue of human memory and seems undesirable in online information retrieval systems (see also Caid et al., 1995). A concrete example will help illustrate the problem.

A textual database can be represented by means of a term-by-document matrix. The database in this example consists of the titles of 9 Bellcore Technical Memoranda. There are two classes of documents -5 about human-computer interaction and 4 about graph theory.

Title Database:

The term-by-document matrix corresponding to this database is shown in Table 1 for terms occurring in more than one document. The individual cell entries represent the frequency with which a term occurs in a document. In many information retrieval applications these frequencies are transformed to reflect the ability of words to discriminate among documents. Terms that are very discriminating are given high weights and undiscriminating terms are given low weights. Note also the large number of 0 entries in the matrix-most words do not occur in most documents, and most documents do not contain most words