[Colloquium] Fitz/MS Presentation/Jun 12, 2017

[Margaret Jaffey via Colloquium]

This is an announcement of Stephen Fitz's MS Presentation.

------------------------------------------------------------------------------
Date:  Monday, June 12, 2017

Time:  2:00 PM

Place:  Ryerson 277

M.S. Candidate:  Stephen Fitz

M.S. Paper Title: Context Dependent Natural Language Understanding
with Deep Neural Networks

Abstract:
This thesis presents two of the three research projects that I’m
involved with. The common theme is the design of neural architectures
for natural language understanding with specific goals in mind. The
natural language input is interpreted in the context of the task, and
in some cases external inputs such as environment observations. The
first project, developed out of my internships in Tokyo during the
past two summers. Japanese ministry of education has been investing
into applications of artificial intelligence to education, and I was
tasked to explore the possibility of advancing the state of the art in
automated essay scoring using modern methods. In the United States,
Educational Testing Service (ETS) has lead the efforts to automate
essay grading based on a variety of techniques from NLP and AI, which
resulted in the e-rator engine that automatically identifies features
related to writing proficiency in student essays so they can be used
for scoring and feedback of exams such as Test of English as a Foreign
Language (TOEFL). However, the e-rator methodology is based on a
mixture of heuristics, traditional AI, and statistical machine
learning, which limits its applicability to settings outside of the
exams it was designed for. Since, the development of the e-rator
engine, a shockwave of discoveries - the deep learning revolution of
recent years - has swept over the AI community. After initial success
in computer vision, we saw an increasing success of recurrent neural
networks such as LSTM, and more recently also convolutional
architectures, in natural language processing. In this project, I
develop a deep neural network based automated essay scoring framework
using an array of methods inspired by the latest advances in deep
learning.

Second project presented here concerns human-machine interaction using
natural language. Following recent trends in AI, we decide again to
follow the deep learning path towards robot intelligence. In
particular, we study the problem of instruction following by an
autonomous vehicle (e.g. a self-driving car) navigating the streets of
downtown Chicago. Prior approaches to this problem required humans to
express themselves in artificially constrained domain specific
languages. Here we approach human-machine communication in analogy to
communication between humans using different natural languages - i.e.
we phrase it as an instance of translation. We develop a deep
recurrent network, Long-Short Term Memory (LSTM) based architecture.
This system interprets the natural language commands in the context of
another sequence of perceptual inputs coming from environment
observations. The main obstacle to deep learning in robotics is
scarcity of training data. We develop an agent-based path generator
using the Google Street View API, which is used to produce a large
dataset of routes containing geolocation data together with images and
environment metadata, aligned with natural language instructions
describing navigation directions given to the vehicle.

Finally, we discuss future directions for research inspired through
our experience gained while working on these projects. These projects
can be grouped under the label of deep learning methods for natural
language problems. Most early efforts in this field were directed
towards more practical NLP, and engineering problems (e.g. named
entity recognition, vector embeddings of text, sentiment analysis,
compression, text mining, question answering, discourse
classification). We believe that there is unexplored potential for
these and other newly emerged deep models to explore more fundamental
problems at the intersection of linguistics and computer science. Of
particular interest to us, are tasks such as morphological analysis
for languages with interesting inflectional paradigms, the development
novel neural architectures tailored to answering more fundamental
questions about language, as well as visualization and mathematical
analysis (e.g. through means of homological algebra) of such models in
order to shed light on what they are learning when trained on language
data. Can neural networks provide new insights into the nature of
linguistic phenomena, above what can be extracted through statistical
machine learning approaches?

Stephen's advisor is Prof. John Goldsmith

Login to the Computer Science Department website for details:
 https://www.cs.uchicago.edu/phd/ms_announcements#stephenf

=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=
Margaret P. Jaffey            margaret at cs.uchicago.edu
Department of Computer Science
Student Support Rep (Ry 156)               (773) 702-6011
The University of Chicago      http://www.cs.uchicago.edu
=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=