Deep Assertion discovery using word embeddings

In recent years, there has been explosive growth in the amount of biomedical data (e.g., publications, notes from EHRs, clinical trial results), with the majority being unstructured data. As the volume of data is increasing faster, the demand for extracting knowledge from unstructured data increases in biomedical research. Many information extraction techniques like OpenIE, Ollie, ReVerb have been proposed to extract binary relations from plain text. The result of these techniques is triplets containing two argument phrases from the input text and a relation between these arguments which are a trade-off between relevance and completeness. These existing IE frameworks are unable to deal with big data. In this thesis, we propose an assertion framework that dynamically refines the triplets to form factual tuples called ‘assertions’ from multiple free-text sources. The framework is built using a pipeline approach to discover the significant features (words) and map them to the relation tuples extracted using the open information extraction technique. We have implemented three different feature discovery techniques – using the numerical statistic like TF-IDF (Term Frequency-Inverse Document Frequency), using word embedding approaches like Word2Vec in Deep Learning frameworks. We have shown that the assertions discovered using our proposed framework can be effectively used for the topic-based classification of the biomedical dataset containing abstracts for three categories such as obesity, heart disease and diabetes from NCBI PubMed and a benchmark dataset, 20 Newsgroups. The knowledge discovery from data in the proposed framework was verified through the classification using a five-layered Convolutional Neural Network. We show that the classification accuracy with assertions derived from the data improves the accuracy of the existing approach like deep belief networks by more than 10% on the benchmark dataset like 20 Newsgroups.

Name thesis:degree_name

M.S.

Level thesis:degree_level

Masters

Discipline thesis:degree_discipline

Computer Science (UMKC)

Grantor dc:publisher

University of Missouri -- Kansas City

Year dc:date.issued

2018

Author dc:creator

• Nagabhushan, Megha

Advisor dc:contributor.advisor

• Lee, Yugyung, 1960-

Language dc:language.iso

en_US

Handle dc:identifier.uri

https://hdl.handle.net/10355/67047

OAI identifier oai:identifier

oai:mospace.umsystem.edu:10355/67047