Date of Award

August 2017

Degree Type

Dissertation

Degree Name

Doctor of Philosophy

Department

Biomedical and Health Informatics

First Advisor

Rohit J. Kate

Committee Members

Rohit J. Kate, Mary Shimoyama, Susan McRoy, Jake Luo

Keywords

Classification, Decision Tree, Named Entity Recognition, Natural Language Processing, Text Mining, Unsupervised

Abstract

Named entity recognition (NER) from text is an important task for several applications, including in the biomedical domain. Supervised machine learning based systems have been the most successful on NER task, however, they require correct annotations in large quantities for training. Annotating text manually is very labor intensive and also needs domain expertise. The purpose of this research is to reduce human annotation effort and to decrease cost of annotation for building NER systems in the biomedical domain. The method developed in this work is based on leveraging the availability of resources like UMLS (Unified Medical Language System), that contain a list of biomedical entities and a large unannotated corpus to build an unsupervised NER system that does not require any manual annotations.

The method that we developed in this research has two phases. In the first phase, a biomedical corpus is automatically annotated with some named entities using UMLS through unambiguous exact matching which we call weakly-labeled data. In this data, positive examples are the entities in the text that exactly match in UMLS and have only one semantic type which belongs to the desired entity class to be extracted (for example, diseases and disorders). Negative examples are the entities in the text that exactly match in UMLS but are of semantic types other than those that belong to the desired entity class. These examples are then used to train a machine learning classifier using features that represent the contexts in which they appeared in the text. The trained classifier is applied back to the text to gather more examples iteratively through the process of self-training. The trained classifier is then capable of classifying mentions in an unseen text as of the desired entity class or not from the contexts in which they appear.

Although the trained named entity detector is good at detecting the presence of entities of the desired class in text, it cannot determine their correct boundaries. In the second phase of our method, called “Boundary Expansion”, the correct boundaries of the entities are determined. This method is based on a novel idea that utilizes machine learning and UMLS. Training examples for boundary expansion are gathered directly from UMLS and do not require any manual annotations. We also developed a new WordNet based approach for boundary expansion.

Our developed method was evaluated on three datasets - SemEval 2014 Task 7 dataset that has diseases and disorders as the desired entity class, GENIA dataset that has proteins, DNAs, RNAs, cell types, and cell lines as the desired entity classes, and i2b2 dataset that has problems, tests, and treatments as the desired entity classes. Our method performed well and obtained performance close to supervised methods on the SemEval dataset. On the other datasets, it outperformed an existing unsupervised method on most entity classes. Availability of a list of entity names with their semantic types and a large unannotated corpus are the only requirements of our method to work well. Given these, our method generalizes across different types of entities and different types of biomedical text. Being unsupervised, the method can be easily applied to new NER tasks without needing costly annotations.

Download

Share

COinS