Transfer learning for low-resource natural language analysis
Author(s)
Zhang, Yuan, Ph. D. Massachusetts Institute of Technology
DownloadFull printable version (3.288Mb)
Other Contributors
Massachusetts Institute of Technology. Department of Electrical Engineering and Computer Science.
Advisor
Regina Barzilay.
Terms of use
Metadata
Show full item recordAbstract
Expressive machine learning models such as deep neural networks are highly effective when they can be trained with large amounts of in-domain labeled training data. While such annotations may not be readily available for the target task, it is often possible to find labeled data for another related task. The goal of this thesis is to develop novel transfer learning techniques that can effectively leverage annotations in source tasks to improve performance of the target low-resource task. In particular, we focus on two transfer learning scenarios: (1) transfer across languages and (2) transfer across tasks or domains in the same language. In multilingual transfer, we tackle challenges from two perspectives. First, we show that linguistic prior knowledge can be utilized to guide syntactic parsing with little human intervention, by using a hierarchical low-rank tensor method. In both unsupervised and semi-supervised transfer scenarios, this method consistently outperforms state-of-the-art multilingual transfer parsers and the traditional tensor model across more than ten languages. Second, we study lexical-level multilingual transfer in low-resource settings. We demonstrate that only a few (e.g., ten) word translation pairs suffice for an accurate transfer for part-of-speech (POS) tagging. Averaged across six languages, our approach achieves a 37.5% improvement over the monolingual top-performing method when using a comparable amount of supervision. In the second monolingual transfer scenario, we propose an aspect-augmented adversarial network that allows aspect transfer over the same domain. We use this method to transfer across different aspects in the same pathology reports, where traditional domain adaptation approaches commonly fail. Experimental results demonstrate that our approach outperforms different baselines and model variants, yielding a 24% gain on this pathology dataset.
Description
Thesis: Ph. D., Massachusetts Institute of Technology, Department of Electrical Engineering and Computer Science, 2017. This electronic version was submitted by the student author. The certified thesis is available in the Institute Archives and Special Collections. Cataloged from student-submitted PDF version of thesis. Includes bibliographical references (pages 131-142).
Date issued
2017Department
Massachusetts Institute of Technology. Department of Electrical Engineering and Computer SciencePublisher
Massachusetts Institute of Technology
Keywords
Electrical Engineering and Computer Science.