FOA Home
Nothing is more frustrating than spending many, many hours on a
technical problem, unless it is then finding out that someone else had
{previously solved} this same problem! One of the motivations shared by
many working on search engine technology is the hope that we can reduce
the number of times the same wheel is reinvented.
D.R. Swanson has
concentrated on the scientific and medical literature, and viewed it as
``a potential source of new knowledge.'' [Swanson90] [Swanson97] . Each new medical report
contains new knowledge about some particular disease or treatment. But
Swanson has gone the next step, to imagine what modes of inference might
be most appropriate across a {\em network} of such papers each
describing potential causal relations: Each scientific article
contributes to a web of logical connections that interlace the
literature of science. Some of these connections are made explicit
through references from one article to another, citatipns that reflect,
among other things, authors' perceptions of how their own work is
related to that of others and how it fits into the scheme of existing
knowledge. However, there may exist many implicit logical interarticle
connections that are unintended and not marked by citations; such
implicit links are the focus of this paper. (The word ``logical'' here
is used informally; a ``logical connection'' is formed by statements
that are related by any process of scientific reasoning or argument.)
Scientific articles can be seen as clustering into more or less
independent sets or ``literatures." Within each set, common probiems are
addressed, common arguments are advanced, and articles ``interact'' by
citing one another. Distinct literatures that are essentially unrelated
are in general "noninteractive" in that they do not cite or refer
directly to each other, have no articles in common, and are not cited
together by other articles. On the other hand, if two literatures are
linked by arguments that they respectively put forward -- that is, are
``logically'' related or connected -- one would expect them to cite each
other. If they do not, then the logical connections between them would
be of great interest, for such connections may be unintended, unnoticed,
and unknown-therefore potential sources of new knowledge.... The number
of possible pairs of literatures (that might be connected) increases
very nearly as the square of the number of literatures; the number of
possible connections increases at even a greater rate if triple or
higher combinations are taken into account rather than just pairs. From
this perspective, the significance of the "information explosion" may
lie not in an explosion of quantity per se, but in an incalculably
greater combinatorial explosion of unnoticed {\em logical} connections.
(p. 29,35)
Swanson's first and most well-known example of new knowledge
discovered in this fashion identified fish oil as a treatment for
Raynaud's syndrome, a circulatory problem resulting in poor blood supply
to the extremities. Figure (figure) shows a second example of
Swanson's method. Beginning with a syndrome like migraine headaches,
Swanson searches the literature for features mentioned in the context of
migraines which also have been mentioned in a second, disjoint,
``mutually oblivious'' literature. Working backwards from the syndrome
to be explained, a query is formed against a medical corpus like
MEDLINE. The resulting set of (in this case 63) documents is then
analyzed and clustered into related sets; it is in this clustering that
Swanson's {\em manual}, intellectual effort is most obvious. The
clusters are then used to suggest additional, new queries. Finally, the
single common ``cause'' of MAGNESIUM is identified. Fig.
(FOAref) also shows an attempt to classify the relationship
between common phrases in the retrieved literature and the relation to
the syndrome. Further, the relationship between these same phrases, or
lexical variations on them! and the common cause magnesium is also
forced into structured relationships.
It is hard to imagine a more
exciting prospect for the analysis of all literature. The identification
of such ``undiscovered public knowledge'' is almost certainly possible
in many other situations. The question becomes how we might
algorithmically search for all of them. Note especially the liberties
taken in interpreting the literature and phrases used consistently
within it as they have been transformed into the {structured relations}
of Fig. (FOAref) . These relations are meant to suggest more
formal and powerful modes of inference between causes and effects
mentioned within each paper. Some of the arrowheads show suggest causal
relationships; some relationships (e.g., {\tt associated\_with}) are
neutral with respect to correlation versus causation; others ({\tt
type\_of}) suggest hierarchic relations between classes.
Swanson is
well-aware of these difficulties: ... [the] form and structure of
logically connected arguments are in general recognizable by scientists
irrespective of their specialty, a point that may have implications for
research on futuristic, more fully automated systems. However, the
simple structure of the syllogistic model does not in many respects
reflect the depth or range of actual problems that would be encountered
if one tried to build a database of logical connections\ldots. The
objective, moreover, is not simply to draw mechanistic logical
inferences, but rather to deter- mine whether certain plausible
connections or hy- potheses appear to be worth testing. Most articles
harbor, either explicitly or implicitly, an enormous number of logical
connections. Which connections are relevant and important can be de-
termined only in the light of a specific context, hy- pothesis, problem,
or question; but such contexts and questions are always changing. The
degree to which one can hope to encode logical connections in any form
suitable for all future uses may therefore be quite limited. (p. 35)
Recognizing that even if fully automatic discovery of new facts is
currently too hard, Gordon and Lindsey have investigated forms of
``discovery support'' [Gordon96] .
Gordon and Dumais have also explored the use of LSI techniques (cf.
Section §5.2.3 ) as part of the
literature-based discovery process [Gordon98] . The formulization of SVD,
eigenstructure analysis of the relationship between a query and
documents (step 2) and then the analysis of these intermediate
literatures and ultimate causes (step 3) is an important extension
beyond the manual investigations originally proposed by Swanson.
Top of Page
Discovering latent knowledge within a corpus