Record Linkage, Text-Mining, and e-Discovery
Paul Thompson
Dartmouth College
A primary focus of the TREC Legal Track is to determine whether more advanced
information retrieval technologies can better support the e-discovery process than
Boolean retrieval, which is the predominant form of information retrieval in use in the
legal domain. The ICAIL 2007 DESI workshop is exploring whether artificial
intelligence technologies can support e-discovery better not only than Boolean, but also
better than statistical ranked retrieval. This paper considers the problem of e-discovery in
a broader context. First, Dartmouth’s Language Technologies Group plans to build a
retrieval system incorporating advanced information extraction technology, in particular
cross-document co-reference of named entities extracted from text. Similar technology
has also been considered in the domain of national security, e.g., see the 2005 SIAM
International Conference on Data Mining Workshop on Link Analysis Counterterrorism
and Security Analysis (SIAM 2005 ), where workshop participants were encouraged to
use the Enron e-mail dataset as a surrogate for a counterterrorism database.
One technology that has been used for link analysis was provided in a commercial
product, Non-Obvious Relationship Analysis (NORA) of SRD. This product was
developed originally to detect fraudulent collusion in the gaming industry in Nevada.
SRD was since acquired by IBM. In a 2006 presentation to the National Academy of
Sciences Jeff Jonas, formerly of SRD, now at IBM, described a way to perform record
linkage with encrypted data, thus preserving privacy (Jonas 2006). Other users of
NORA, however, claim that record-linkage effectiveness would be degraded were
analysis to be done with encrypted data. Dartmouth’s approach to record linkage and text
mining for e-discovery will also include process query system technology to model
dynamic relationships among entities in text (Cybenko and Berk 2007, Patman and
Thompson 2003, Thompson 2005)
In future e-discovery, when the new retrieval methods pioneered by the TREC Legal
Track, the ICAIL community, or the commercial e-discovery sector have matured,
protecting privacy in e-discovery may be even more important than it is now. For
example, question answering, text mining, or multimedia data mining technologies may
be used to sample documents, broadly defined, from the total information repository of
an organization.
Our research proposes to compare the performance of a retrieval system augmented by
record linkage and text mining capabilities on the TREC Legal Track corpus under two
conditions. The same retrieval system will be used first with unencrypted data and then
with encrypted data. The degradation in performance, if any, will be measured.
References
SIAM. 2005. SIAM International Conference on Data Mining Workshop on Link
Analysis Counterterrorism and Security Analysis
http://www.cs.queensu.ca/~skill/proceedings/
Cybenko, G. and Berk, V. 2007. “Process Query Systems” IEEE Computer, January
2007, p. 62-70
Jonas, J. 2006. “The Information Sharing Paradox And Thoughts on Maximizing
Discovery while Minimizing Disclosure” presentation to the National Academies Project
on Technical and Privacy Dimensions of Information for Terrorism Prevention and Other
National Goals
Patman, F. and Thompson, P. “Names: A New Frontier in Text Mining” NSF / NIJ
Symposium on Intelligence and Security Informatics, Lecture Notes in Computer Science,
vol. 2665, Berlin: Springer-Verlag, June 1-3, 2003, Tucson, Arizona, p. 27-38
Thompson, P. “Text Mining, Names, and Security” Journal of Database Management,
vol. 16, no. 1, January-March, 2005, special issue on database technology for enhancing
national security, p. 54-59