Off-campus UMass Amherst users: To download campus access dissertations, please use the following link to log into our proxy server with your UMass Amherst user name and password.

Non-UMass Amherst users: Please talk to your librarian about requesting this dissertation through interlibrary loan.

Dissertations that have an embargo placed on them will not be available to anyone until the embargo expires.

Author ORCID Identifier

https://orcid.org/0000-0003-2798-0726

Document Type

Open Access Dissertation

Degree Name

Doctor of Philosophy (PhD)

Degree Program

Computer Science

Year Degree Awarded

2019

Month Degree Awarded

September

First Advisor

Andrew McCallum

Subject Categories

Artificial Intelligence and Robotics

Abstract

Natural language processing (NLP) has come of age. For example, semantic role labeling (SRL), which automatically annotates sentences with a labeled graph representing who did what to whom, has in the past ten years seen nearly 40% reduction in error, bringing it to useful accuracy. As a result, a myriad of practitioners now want to deploy NLP systems on billions of documents across many domains. However, state-of-the-art NLP systems are typically not optimized for cross-domain robustness nor computational efficiency. In this dissertation I develop machine learning methods to facilitate fast and robust inference across many common NLP tasks.

First, I describe paired learning and inference algorithms for dynamic feature selection which accelerate inference in linear classifiers, the heart of the fastest NLP models, by 5-10 times. I then present iterated dilated convolutional neural networks (ID-CNNs), a distinct combination of network structure, parameter sharing and training procedures that increase inference speed by 14-20 times with accuracy matching bidirectional LSTMs, the most accurate models for NLP sequence labeling. Finally, I describe linguistically-informed self-attention (LISA), a neural network model that combines multi-head self-attention with multi-task learning to facilitate improved generalization to new domains. We show that incorporating linguistic structure in this way leads to substantial improvements over the previous state-of-the-art (syntax-free) neural network models for SRL, especially when evaluating out-of-domain. I conclude with a brief discussion of potential future directions stemming from my thesis work.

Creative Commons License

Creative Commons Attribution 4.0 License
This work is licensed under a Creative Commons Attribution 4.0 License.

Share

COinS