[Colloquium] Research at TTIC: Mohit Bansal, TTIC Research Ass't Professor

[Dawn Ellis]

When:     Friday, February 7th at noon.

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

Speaker:  Mohit Bansal, TTIC Research Ass't Professor

Title:        Syntactic Parsing with Word Embeddings

Abstract:

Unsupervised word representations, derived from unlabeled text data, have
proven very useful in many supervised NLP tasks. Most such representations
fall into one of two categories: discrete representations which consist of
memberships in a hard clustering of words, and continuous representations
(or distributed representations or embeddings) which consist of dense,
real-valued vectors for each word, typically induced via neural language
models or spectral methods.

Dependency parsing is the task of predicting the syntactic structure of a
natural language sentence in the form of binary, asymmetric dependency
relations between the words of the sentence. Previous work has reported
substantial gains in dependency parsing by adding features based on
discrete word representations via Brown clusters derived from large,
unlabeled corpora. I will present our recent work on the use of continuous
word embeddings for dependency parsing, such as representations induced by
neural language models. We compare several types of such word
representations as well as Brown clusters, and evaluate them on syntactic
and semantic word-similarity, and as features in in-domain news (WSJ)
parsing and open-domain Web parsing.

In work-in-progress, I will discuss syntactically-trained embeddings and
multi-view embeddings for syntactic parsing.

This is joint work with Kevin Gimpel and Karen Livescu.

-- 
*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/20140203/b68a5d3e/attachment-0001.htm