Xerox Research Centre Europe, located in Grenoble, is offering the
following internship for Spring 2015:

 

-------------------------------------------------------------------

Internship: Modeling next sentence in a dialog using vector space
models

-------------------------------------------------------------------

 

See: http://www.xrce.xerox.com/About-XRCE/Internships/Modeling-next-sentence-in-a-dialog-using-vector-space-models

 

Contacts :

                Dymetman, Marc             marc.dymetman@xrce.xerox.com  

                Venkatapathy, Sriram    sriram.venkatapathy@xrce.xerox.com

 

Please mention "Modeling next sentence in a dialog using vector space
models" in your subject line.

 

Duration: 4.5 months

Start Date: April - June 2015

 

Given the context of an observed dialog history up to a certain point,
the goal of this internship will be to develop predictive models for
the next utterance. A good model of the next utterance should
substantially reduce its ``perplexity'' with respect to a baseline
language model. Based on a large collection of available chats in the
customer-care domain, the feasibility of learning such models will be
explored. The perplexity score will be used to evaluate the
performance of various models that will be examined. The internship
will also involve using such models for the task of actually
predicting the next utterance, up to some limited editing actions. The
idea will be to retrieve the top candidates for the next sentence from
all the sentences that have ever been uttered. One of the criteria for
retrieval will be the perplexity of these sentences relative to the
learnt model.

 

For modelling, we will focus on models that represent the dialog
history as a real vector. Such representations have been extensively
used in recent years for encoding the semantics of words and
sentences. Several ideas will be explored using frameworks such as
Topic modelling and Deep Neural Networks.

 

The ideal candidate is a Computer Science student at the Master or
(preferably) PhD level, with a background in Machine Learning and
NLP. Strong programming skills are a requirement (Python, C++,
Java...). Conference publication of the work will be strongly
encouraged.