Authors

Ritwik Banerjee

Type

Text

Type

Dissertation

Advisor

Choi, Yejin | Ramakrishnan, I. V. | Fodor, Paul | Balasubramanian, Niranjan | Davulcu, Hasan.

Date

2015-12-01

Keywords

information extraction, natural language processing | Computer science

Department

Department of Computer Science.

Language

en_US

Source

This work is sponsored by the Stony Brook University Graduate School in compliance with the requirements for completion of degree.

Identifier

http://hdl.handle.net/11401/77269

Publisher

The Graduate School, Stony Brook University: Stony Brook, NY.

Format

application/pdf

Abstract

A wealth of knowledge in the biomedical domain is available in unstructured or semi-structured data repositories as natural language narratives. Much of this knowledge can provide immediate and tangible benefits in patient welfare and the healthcare industry. Extracting relevant knowledge from these natural language sources and providing them as structured information suitable for immediate real-time consumption in clinical settings is, however, a manual process restricted to human domain experts. As a result, it is expensive and time-consuming. A very real consequence of this is that the journey made by medical " knowledge nuggets" from research publications to patient care settings like hospitals often take several years. Even so, the knowledge still gets presented to clinicians in natural language -- unsuitable for machine consumption, and an impediment to the pace of work often demanded of clinicians (e.g. | in emergency rooms). Automatic extraction of this knowledge is a challenging task. Biomedical research literature is replete with language constructs that are highly specific to not just the domain, but internal sub-domains. The linguistic semantics used in discussions of, say, diabetes, are very different from the semantics used to discuss diseases like malaria that are caused by external agents. Moreover, being research literature, authors typically write for readers with a fair amount of encyclopaedic domain knowledge. Consequently, important information can often only be gleaned by identifying causal relations that are implicit. Standard information extraction methods that depend on identifying causality in text usually require explicit discourse connectives like " because" , " since" , etc. Additionally, they manage to extract only those relations that are expressed within the span of a single sentence. This proposal presents a novel relation learning methodology for biomedical natural language that is able to infer relations where (a) the relation is implicit, and (b) the related entities do not co-occur within the span of a single sentence. We show that our technique outperforms a sentence-level supervised classification approach. Further, as a human-in-the-loop (HITL) model, it is capable of augmenting biomedical knowledge bases quickly and accurately. Finally, we contribute two novel applications that demonstrate the use of such relational knowledge in providing real-time clinical decision support. | 75 pages

Share

COinS
 
 

To view the content in your browser, please download Adobe Reader or, alternately,
you may Download the file to your hard drive.

NOTE: The latest versions of Adobe Reader do not support viewing PDF files within Firefox on Mac OS and if you are using a modern (Intel) Mac, there is no official plugin for viewing PDF files within the browser window.