Mining the Bibliome - Translational Informatics: Realizing the Promise of Knowledge-Driven Healthcare - page 78

IBM [ 19 ], which harnesses available knowledge including that are encoded into
natural language formats. It is important to acknowledge the diffi culty in NLU and
the interpretation of human discourse - it is not uncommon for meaning to be con-
veyed through idioms or indirect references. Other challenges are faced by NLU
systems, such as resolving abbreviations, connecting concepts across statements;
and disambiguation of identifi ed entities, form the inspiration for research. To initi-
ate the process of knowledge discovery, complete NLU functionality that equates to
human understanding may not be required. The ability for entity recognition and
semantic reconciliation of identifi ed concepts has evolved to the point that a number
of publicly available systems can be used with confi dence within research environ-
ments. Two commonly used systems in biomedicine are the MetaMap system from
the National Library of Medicine [ 20 ] and Annotator from the National Center for
Biomedical Ontology (NCBO) [ 21 ].
The vast majority of bibliome mining approaches and resources in biomedicine
are geared towards researchers. There is some energy in developing “real-time deci-
sion support” that would provide some active support for clinicians; however, most
decision support applications are based on passive decision support. In contrast to
active decision support systems, where important knowledge inferences are made in
real time to clinicians through interactive interfaces, passive decision support
systems are based on searching already curated (either manually or through the use
of bibliome mining algorithms). An increasingly popular exemplar of this type of
passive decision support is the “infobutton,” which is increasing being integrated
into modern electronic health record systems [ 22 ]. Infobuttons work through pro-
viding a guided interface to information resources, such as MEDLINE for biomedi-
cal literature. In addition to a number of systems that have been designed by
dedicated researchers for analyzing specifi c types of natural language (e.g . , for ana-
lyzing clinical texts, there are a number of well-described systems like MedLEE
[ 23 ], MPLUS/ONYX [ 24 , 25 ], or MedSynDiKATe [ 26 ]) there is a continued need
to develop NLU systems to extract usable information from the range of natural
language sources (including those that are more general in nature, like the ARC
system that was designed to be 90 % good for 90 % of information extraction tasks
[ 27 ]). Historically, NLU systems were designed as specifi c solutions (commonly
written in logic programming languages such as Prolog). In recent years, common
programming frameworks have facilitated the ability to develop NLU systems that
can be more community driven. The two most prevalent systems are the General
Architecture of Text Engineering (GATE [ 28 ]) and Unstructured Information
Management Architecture (UIMA [ 29 ]). By being community driven, it is possible
to combine features and functionality into new systems that can meet specifi c infor-
mation extraction needs as well as design new techniques that can be shared and
enhanced by the community. Active decision support systems that leverage bibli-
ome mining techniques are often termed “question-answer” tools, where a clinician
(or anyone with a biomedical question) can present a question in natural language
and then the response is based on bibliome mining of all available corpora [ 30 ]. The
most well-known examples of these types of question-answer system includes the
aforementioned IBM Watson as well as the publicly accessible WolframAlpha [ 31 ]