Events Calendar

NL Seminar-Multitask Word Alignment with Random-Walk Regularizers

Speaker: Tomer Levinboim, Notre Dame University

Talk Title: Multitask Word Alignment with Random-Walk Regularizers

Series: Natural Language Seminar

Abstract: Suppose we translate a word from English to French and back. Should we get the original English word? That is, is translation invertible? Alternatively, suppose we translate an English word e to Spanish and then from Spanish to French, obtaining a word f. Should e-f be a valid entry in an English-French dictionary? That is, is translation transitive? Intuitively, if translation is done carefully, we expect to answer both these questions with "Yes, with high probability". In this talk, I will discuss how to formulate our intuition about invertibility/transitivity with random-walks, using translation probability distributions. I will then present two random-walk based regularization techniques that we recently used in a multitask word alignment setting: (1) Model Invertibility Regularization (MIR) - a concave regularizer for bi-directional models which can be applied even without parallel data. (2) Triangulation based Dirichlet prior - a method that capitalizes on parallel data with a pivot language, to construct and learn better translation priors. This talk is based on joint work with Prof. David Chiang (ND) and Dr. Ashish Vaswani (ISI).




Biography: Tomer Levinboim is a PhD student at the University of Notre Dame, working with Prof. David Chiang on developing machine learning techniques for improving machine translation and NLP of low resource languages. He is generously hosted by Kevin Knight at USC/ISI.

Host: Nima Pourdamghani and Kevin Knight

More Info: http://nlg.isi.edu/nl-seminar/

Location: Information Science Institute (ISI) - 6th Flr Conf Rm # 689, Marina Del Rey

Audiences: Everyone Is Invited

Contact: Peter Zamar

Event Link: http://nlg.isi.edu/nl-seminar/