untitled

146 6 Information Retrieval

Summary

• Citations (such as hypertext links) can be used to rank documents relevant
  to a query according to various criteria:
  
  
  1. Authoritativeness
  
  
  2. Central source
  
  
  3. Popularity
  
  
  
  


• Citation ranking improves both selectivity and coverage.


• However, citation ranking has a number of weaknesses:
  
  
  1. Current systems are based on matching words in the query with words
     in documents and do not consider the meaning of the words.
  
  
  2. Only the principal eigenvector is used, so smaller communities will be
     masked by larger ones.
  
  
  
  

6.5 Vector Space Retrieval of Knowledge Representations

One of the main assumptions of the vector space model is that documents

are composed of collections of terms. While some systems attempt to take

advantage of correlations between terms, such correlations are difficult to

determine accurately, and the number of correlations that must be computed

is huge. In any case, the terms are still disjoint from one another. Knowledge

representations change this situation. Terms can now be complex concept

combinations that are built from simpler terms. Thus a term like “flu vac-

cine” contains both “flu” and “vaccine” as well as the complex relationship

between these two concepts which expresses the effect of the vaccine on the

influenza virus as well as the the derivation of the vaccine from the virus and

in response to it. In the UMLS, all three of these are concepts, and they are

related to one another.

To see how natural, as well as how subtle, concept combinations can be,

try juxtaposing two commonly used terms in different orders. For example,

“test drug” and “drug test.” Although these two have completely different

meanings, most search engines give essentially the same answer for both.

Indeed, “test drug” can be interpreted in two ways depending on whether

“test” is a verb or adjective. The term “drug test” also has several meanings.

As an exercise, try some other pairs of terms to see how many meanings