111-27

that, on overlap, non-relevant documents are longer than relevant, the length

being c~nsiderably above the average.   Similar parameters are calculated

for the cosine ordering in Figures l~ and 15, and very similar lengths are

then obtained for relevant and non-relevant documents in this case.

        Since the analysis shows that non-relevant documents with strong

matches are longer than average, it is now obvious that cosine effectively

lowers the ranks of these documents, and thus provides a better retrieval

performance than overlap.  Although it is certainly the case that non-relevant

documents with weak matches must be shorter than average, it seems that their

low match (sometimes as low as zero) is never sufficient to increase their

rank by any significant amount on overlap; it is the strongly matched non-

relevant only that are responsible for the superiority of the cosine correlation.

        This phenomenon is probably caused by the fact that not all non-

relevant documents have an equal probability of resulting in spurious matches;

as seems logical, the probability of spurious matches is greater in larger

documents.  Spurious matches result from spurious concept combinations, which

arise because no judgments of importance are made to discriminate between

request concepts; that is, any combination of, say, three concepts (out of

six in a request) is assumed to be as important as any other.  An example of

this is given in ~igure 14, where both a non-relevant and a relevant document

match in three'out of the six concepts; the data of Figure 16 show, however,

that the non-relevant match on words such as ??~~~gIl,  "report" and

`I

measurement" which turn out to be spurious.    Such spurious matches are more
likely to occur for long non-relevant documents than for short ones.   The

logical search formulations used in post-coordinate manual systems would

eliminate many such false matches; some success in this direction can be

achieved without manual search form~~ation by use of weighting methods, to ke

described next.