[Colloquium] REMINDER: Talks at TTIC: Kevin Gimpel, TTIC

[Dawn Ellis]

When:     Monday, March 16, 2015 at 11am

Where:    TTIC, 6045 S Kenwood Avenue, 5th Floor, Room 526

Who:       Kevin Gimpel, TTIC

Title:       Learning Criteria for Natural Language Processing

Abstract:

Natural language processing (NLP) tasks exhibit a range of characteristics
that complicate the application of machine learning. Outputs are often
richly-structured, sometimes with structured latent variables, and are
scored with complex evaluation metrics. Gold standard annotations often
have multiple legitimate possibilities, especially for linguistic
generation tasks like translation, summarization, and image caption
generation. While unlabeled linguistic data is abundant, it is unclear how
to use it to improve performance on a particular task. And while annotated
data is available for many tasks, it is often not in the language or domain
of interest.

I will discuss my work on designing objective functions to address (and
exploit) these characteristics of NLP tasks. I'll first discuss
softmax-margin, an objective function for structured prediction that
combines latent variables with task-specific evaluation metrics. I will
present results for sequence labeling and mention follow-up work done by
others for other structured NLP tasks. I will then discuss a family of
structured ramp losses which have attractive theoretical properties and
naturally handle the issue of multiple correct outputs. A novel
variation---hope-fear ramp loss---performs well for machine translation and
has also been found effective for speech recognition and summarization.
Finally, I'll discuss objective functions for weakly-supervised learning
that can leverage unlabeled data as well as annotations in many languages.
Experiments will be reported for unsupervised part-of-speech tagging,
achieving state-of-the-art results.

I'll close with a discussion of next steps, specifically focusing on text
understanding with rich contextual clues. These clues can include the
intent of the author, the socio-temporal context of the text, and the
cognitive inventory of the reader as she reads through the discourse,
including the application of common sense reasoning. I plan to pursue both
traditional symbolic approaches as well as linguistically-motivated neural
architectures; the latter are currently an active area of research in NLP,
and I will speculate briefly on next steps in this direction.

Bio:
Kevin Gimpel is a research assistant professor at TTI-Chicago. He received
his PhD in 2012 from the Language Technologies Institute at Carnegie Mellon
University, where he was an inaugural member of Noah's ARK. His research
focuses on natural language processing, focusing on applications like
machine translation, syntactic analysis of social media, and text-driven
forecasting of real-world events. He also works on machine learning
motivated by NLP, including approximate inference for structure prediction
and learning criteria for supervised and unsupervised learning. He received
a five-year retrospective best paper award for a paper at WMT 2008.

Host:  Karen  Livescu, klivescu at ttic.edu



-- 
*Dawn Ellis*
Administrative Coordinator,
Bookkeeper
773-834-1757
dellis at ttic.edu

TTIC
6045 S. Kenwood Ave.
Chicago, IL. 60637
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mailman.cs.uchicago.edu/pipermail/colloquium/attachments/20150313/419c475b/attachment.htm