The New Jersey Institute of Technology's
Electronic Theses & Dissertations Project

Title:	Transfer learning: bridging the gap between deep learning and domain-specific text mining
Author:	Cheng, Chaoran
View Online:	njit-etd2020-010 (xiii, 76 pages ~ 4.8 MB pdf)
Department:	Department of Computer Science
Degree:	Doctor of Philosophy
Program:	Computer Science
Document Type:	Dissertation
Advisory Committee:	Wei, Zhi (Committee chair) Calvin, James M. (Committee member) Geller, James (Committee member) Basu Roy, Senjuti (Committee member) Guo, Wenge (Committee member)
Date:	2020-05
Keywords:	Machine learning Multimodal text mining Named entity recognition Natural language processing Text data augmentation
Availability:	Unrestricted
Abstract:	Inspired by the success of deep learning techniques in Natural Language Processing (NLP), this dissertation tackles the domain-specific text mining problems for which the generic deep learning approaches would fail. More specifically, the domain-specific problems are: (1) success prediction in crowdfunding, (2) variants identification in biomedical literature, and (3) text data augmentation for domains with low-resources. In the first part, transfer learning in a multimodal perspective is utilized to facilitate solving the project success prediction on the crowdfunding application. Even though the information in a project profile can be of different modalities such as text, images, and metadata, most existing prediction approaches leverage only the text modality. It is promising to utilize the visual images in project profiles to find out how images could contribute to the success prediction. An advanced neural network scheme is designed and evaluated combining information learned from different modalities for project success prediction. In the second part, transfer learning is combined with deep learning techniques to solve genomic variants Named Entity Recognition (NER) problems in biomedical literature. Most of the advanced generic NER algorithms can fail due to the restricted training corpus. However, those generic deep learning algorithms are capable of learning from a canonical corpus, without any effort on feature engineering. This work aims to build an end-to-end deep learning approach to transfer the domain-specific knowledge to those advanced generic NER algorithms, addressing the challenges in low-resource training and requiring neither hand-crafted features nor post-processing rules. For the last part, transfer learning with knowledge distillation and active learning are utilized to solve text augmentation for domains with low-resources. Most of the recent text augmentation methods heavily rely on large external resources. This work is dedicates to solving the text augmentation problem adaptively and consistently with minimal resources for token-level tasks like NER. The solution can also assure the reliability of machine labels for noisy data and can enhance training consistency with noisy labels. All the works are evaluated on different domain-specific benchmarks, respectively. Experimental results demonstrate the effectiveness of those proposed methods. The advantages also indicate promising potential for transfer learning in domain-specific applications.