Text Mining: Opportunities and Barriers

Text Mining:

Opportunities and Barriers

John
McNaught

Deputy Director

National Centre for Text Mining

John.McNaught@manchester.ac.uk


Topics

•
What is text mining? (briefly)

•
What can it offer? (selectively)

•
What are the obstacles? (mostly)


NaCTeM

•
First publicly
-
funded (JISC) national text
mining centre in the world

•
Remit: provide services to research
community

•
Initial focus on biology, then social
sciences, medicine, chemistry, …

•
Processing on a large scale, e.g. for
UKPMC (
Wellcome

T.+17 other funders)

•
www.nactem.ac.uk


What is text mining?

•
Goal: Discover new knowledge from old

•
How:

–
Process very large amounts of text

•
Millions of documents, the more the better

–
Identify and extract information

–
(Link extracted information to already
curated

knowledge)

–
Mine to discover implicit significant associations

–
Flag (unknown) associations for researcher to
investigate further

–
Spin
-
off on the way: render information explicit


From text to new knowledge


What does it offer?

•
Finds unsuspected knowledge

–
E.g. Disease
-
gene associations

•
Enables discoveries human effort could not
achieve (information overload/overlook)

•
Enables better search/navigation of literature

–
Semantic search via extracted semantic metadata

•
Reduces time spent searching

–
15
-
48% of researcher time spent on classic search,
20
-
50% of classic searches unsatisfied

•
E.g. Systematic reviews: months to weeks


What does it offer?

•
Text mining boosts research

–
Makes research possible that would otherwise
be impossible or unfeasible

•
Research drives growth and innovation

•
Research produces more information

•
More information is available for text
mining

•
Text mining boosts research …



Barriers

•
Access to the literature

•
Format issues (tied to next point…)

–
“PDF is evil” (Lynch)

•
Main blocks: copyright and licensing issues

–
<8% of scientific claims found in full article
appear in its abstract (Blake)

–
Abstracts deficient on argumentation,
discussion, methods, background, …

–
Full texts needed to
realise

full benefits of TM



Barriers

•
Need to copy documents to
analyse

them

•
Licences

typically not
favourable

to TM

•
Licences

established on per institution basis

–
Prevents community
-
oriented services

•
Results only for internal use by institutional users

–
Hinders mining over collections of content from
different providers

•
Inconsistency: human can search and
manually
analyse
, but cannot use machine to
do same job on same data already
subscribed to

Barriers

•
Problem even with liberal OA
licences

–
Author attribution required

•
Author attribution in a data mining
environment is impossible/unfeasible

–
Association finding: cannot track positive,
negative, neutral individual author contributions

•
Derived works in a TM environment

–
Every author of every text processed to produce
new derived knowledge may have a claim…


–
Rights clearance thus an effective barrier



Barriers

•
Laudable effort 1: NESLi2
model

licence

(JISC Collections) allows TM

–
Publisher <> single institution

–
But how many publishers retain TM provisions?

–
But cannot display annotations produced by TM
on document itself

•
Laudable effort 2: NPG
licence

for self
-
archived content allows TM

–
But “content must be destroyed when experiment
complete” is vague. So services for community?


Conclusion

•
Copyright and licensing restrictions block full
realisation

of TM benefits

–
Economic savings and potential for growth are
stifled

•
Japan has introduced an information analysis
exception to copyright law

–
National Diet Library (= British Library) has
recently changed its motto to:




“Through knowledge we prosper”

–
Can we say the same in the UK?

Extras


Info=degree of surprise

Finding unknown associations:

reproducing a discovery reported 5 days ago in Nature Medicine

UKPMC
EvidenceFinder

by
NaCTeM
:

Questions generated by deep analysis, with known answers

Click on a question to see relevant extracted evidence

(from OA subset of the archive)