Philip Resnik
Philip Resnik
|
Follow @psresnik
|
|
|
|
The key question I'm exploring: what can the signal available in language use tell us about underlying mental aspects of the speaker/author, such as their ideology, emotional state, or the presence of mental disorders? My work in this area has included topics such as sentiment analysis, persuasion, framing, and "spin", and I'm particularly interested in connections among lexical semantics, surface linguistic expression, and underlying internal state. One area in which I'm excited about applying these ideas is
computational political science. For example, why does my son say "My toy broke"
instead of "I broke my toy"? He's using syntax to package up the
statement about what happened in a way that de-emphasizes semantic
properties such as causation, volition, and change-of-state. (This is
an example of using syntax for "spin", just the same way that Ronald
Reagan did in 1987 when he sidestepped attributing responsibility for
the Iran-contra scandal; remember "Mistakes were made"? Precocious
child.) My
student Stephan
Greene
did a
fascinating dissertation on this topic, and for a
conference-paper-length description see
our 2009
NAACL paper. More recent topics of investigation include
using a sequential topic model to capture topic control in debates and other political conversations (see
this 2013 article in Machine Learning Journal),
extending ideal point models with hierarchical topic modeling to characterize issue framing
(see this ACL 2015 paper looking at Tea Party and Establishment Republicans in the U.S. Congress), and some work in progress on using ideal point models with real-time responses to characterize political reactions in non-political actors. The last of these leverages
work I did with political scientist collaborators on
the React Labs project, a
mobile app for large scale, real-time collection of people's
responses during live events like political debates.
The other application area for these ideas --- and a central focus for my ongoing work --- is applying similar computational models to the identification of linguistic signal related to mental health. After some initial exploration involving social media data and computational modeling related to depression, I've been collaborating with a colleague in the medical school to apply similar ideas to schizophrenia; you, dear reader, are invited to contribute data to this effort here.
I've also been involved in computational social science outside academia, including advising and spearheading development of the sentiment
analysis platform for Converseon Inc., a leading
social media analysis firm (which recently closed a $5M Series A); serving as an advisor to FiscalNote, provider of a real-time legal analytics platform that tracks, analyzes, and forecasts legal and policy data; and founding React Labs to commercialize my work on mobile real-time responses for improved measurement and engagement.
Computational psycholinguistics.
During the past several years, I have been re-engaging more fully with my
longstanding interests in computational psycholinguistics. I'm particularly
interested in the interactions between syntactically mediated compositional processes
and broader context, for which vector space representations (yes, including "deep learning")
offer some interesting modeling tools. Some recent papers related to this line of work include
Ettinger, Phillips, and Resnik, Modeling N400 amplitude using vector space models of word representation (CogSci 2016) and
Ettinger, Resnik, and Carpuat, Retrofitting sense-specific word vectors using parallel text (NAACL 2016). I also remain quite interested in the possibility that ideas from (statistical)
information theory may have a useful role to play in explaining why
language works the way it does. (This is an idea I first began
exploring in my dissertation [ps,
pdf], back in 1993, and in
following years a variety of people like John Hale, Roger Levy, and
Florian Jaeger, among others, have done very interesting work in the
same spirit.) I'm also interested in using Bayesian modeling as a way
to bring linguists here with cognitive modeling interests together
with computational linguists focusing on applications; folks who are interested
in that line of thinking should be talking with Naomi Feldman.
Deep learning. I debated whether to include this here, because frankly I believe there is a ton of hype and many people are excited about so-called "deep learning" (a better term coined by Noah Smith is squash networks) for the wrong reasons. That said, it's practically impossible to get away from this topic, and I'm supervising students who do things with neural network models, so let me say what I do find interesting about this line of work. First, there is a great deal of power in representation learning (unsupervised feature learning) and the sharing of statistical power using subsymbolic representations to improve generalization. Second, deep learning is creating a renewed energy around fundamental scientific questions in computational linguistics that I care about, including the nature of lexical representations, compositional interpretation, and computational models of human sentence processing. Third... Actually, that's about all that comes immediately to mind. :) If you're a prospective student and these approaches interest you, be prepared to show me that you've thought about why they should interest us!
Clinical informatics. Since about 1999 I've been involved in natural language
processing for clinical documentation. I helped start up
CodeRyte, Inc.,
which became the nation's fastest growing provider of NLP solutions in healthcare (showing up in
Deloitte's Technology
Fast 500 and
the Inc. 5000
listings); the company was acquired in April 2012
by 3M
Health Information Systems. I developed major pieces of the core
technology, helped build an excellent language technology team, and I
continue to advise on technology development and strategic
direction.
Somewhere along the way, much to my surprise, I was listed at #82 on the Future Health 100, a list of
"the most creative and influential innovators working in healthcare today"
at healthspottr.com. I don't do a great deal of academic research in this area
largely for reasons having to do with limited access to clinical data, a subject I discuss in my comments on research and social impact under the heading of Electronic Health Records.
Machine translation.
I am no longer actively leading system development in machine translation,
but I collaborate with Marine Carpuat
and students and remain strongly interested in multilingual modeling, e.g.
collaborating with Jordan Boyd-Graber on
multilingual topic modeling in context of the DARPA LORELEI research program. My own approach to MT generally focused on
ways of exploiting parallel corpora and linguistically informed
modeling. Leading the statistical MT efforts at UMD for some dozen years starting in 2003,
I'm pleased to have had a wonderfully productive group.
My then-postdoc David
Chiang (now faculty at Notre Dame) developed
Hiero,
the first syntax-based system to demonstrate performance comparable to
then state-of-the-art statistical phrase-based MT systems (see 2005
NIST MT Evaluation results), and I had the pleasure of working with a number of
students to further improve hierarchical phrase-based translation,
with some innovations including the introduction of lattice
decoding (useful in translation of speech recognition output and also
for text translation of morphologically complex languages),
development of efficient algorithms for using suffix array
representations in hierarchical decoding, use of
English-to-English translation to create artificial reference
translations for use in parameter tuning, the introduction of soft syntactic
constraints based on source language structure, and exploitation of
lattices and forests to represent source language paraphrase and
syntactically driven reordering alternatives.
Crowdsourcing and translation.
Connected with machine translation research, Ben Bederson and I
made an ambitious attempt at progress toward low cost, high quality translation by taking advantage of
monolingual human participants in a computer-assisted translation protocol, in a project we called
"Translation as a Collaborative Process". This blended ideas from machine translation, human computer-interfaces, and
distributed human computation ("crowdsourcing"), and tackled the real-world problem of translating books
in the International Children's Digital Library. We received a
2009 Google Research Award sponsoring
this work, as well as funding from NSF. In September 2009, Ben gave a
Google tech talk about the project which is available on YouTube.
With a follow-up Google Research Award we collaborated with Chris
Callison-Burch to bring his crowdsourcing work and ours together in a framework we called "Translate the World".
Although I'm no longer currently active with this line of research, I remain quite interested in thinking about the
machine translation problem more broadly than just a machine taking a chunk of text in and spitting a chunk of text out:
I think there is a great deal of untapped potential in thinking about the underlying need that MT
serves, which is to produce understanding of language input in another language in ways that are cost-effective, available when you need them, and good enough for the task at hand.
See my on-line list of publications for links
to papers on the above research topics and more.
As of May 2015, major congratulations to Rebecca for transforming Mindwell Psychology Bethesda (originally co-founded with Rachna Varia and Kristi Guadagnoli of Mindwell Psychology in Chantilly) into Rebecca Resnik and Associates, a new practice of which she is founder and director, now fully independent of Mindwell Chantilly.
Click here to arrange a meeting. Otherwise,
in general, the best way to reach me is by e-mail to resnik [AT] umd _DOT_ edu.
Research Interests
Computational social science.
(Why? See my discussion of research and social impact.)
Some additional areas of interest include:
Professional History
.
Professional Activities
Feel free to ask me about the not-yet-written followup: Why I Stopped Working on Machine Translation.
The Rest
Course Information
My regular teaching schedule generally includes a seminar in the fall,
and advanced computational linguistics in the spring.
Current semester
Previous courses
Computational Linguistics Colloquium Series
Other Links
Advice
Handy links
Miscellaneous other
Note that links here may be significantly out of date. This section may be most useful as a window into my odd ideas about what's worth putting up on my web page over the years.
Contact info
Philip Resnik, Professor
Department of Linguistics and Institute for Advanced Computer Studies
1401 Marie Mount Hall
University of Maryland Department phone : (301) 405-7002
College Park, MD 20742 USA Department Fax : (301) 405-7104
http://umiacs.umd.edu/~resnik E-mail: : resnik [AT] umd _DOT_ edu
UMIACS office: AV Williams 3133
Oh, and by the way, my name is not spelled Philip Resnick,
Phillip Resnik, or Phillip Resnick, though this explicit disclaimer
may help people who don't know that find this page!
